Daily 2026 · 07 · 01
TechApplication

Anthropic ships Claude Sonnet 5 with a mid-tier that edges Opus 4.8 on knowledge work

The new default at claude.ai and Claude Code lands at $2 input and $10 output per million tokens through August 31, hits 63.2 percent on SWE-bench Pro and 80.4 percent on Terminal-Bench 2.1, and scores 1,618 on GDPval-AA v2 against Opus 4.8's 1,615. Simon Willison measures the new tokenizer producing roughly 1.4 times more tokens for identical English input, cancelling the headline discount on English workloads.

Anthropic · 2026-06-30 · Source ↗ ·

Anthropic released Claude Sonnet 5 on Tuesday morning and pushed it to the default seat at claude.ai for free and Pro users, inside Claude Code, and across Cursor, VS Code, GitHub Copilot and the raw API. The model landed with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, then reverts to standard rates of $3 input and $15 output in September. Context window: 1,000,000 tokens. Output window: 128,000 tokens. Training cutoff: January 2026. Adaptive thinking is on by default and cannot be turned off through the Claude Code CLI, only through per-organization allowance limits.

The benchmark board is where Anthropic wanted the coverage focused. Sonnet 5 scored 63.2 percent on SWE-bench Pro against Sonnet 4.6's 58.1 and GPT-5.5's 58.6, trailing Opus 4.8 at 69.2. On Terminal-Bench 2.1 the model reached 80.4 percent, up from Sonnet 4.6's 67.0, though GPT-5.5 still leads at 83.4. OSWorld-Verified came in at 78.5 percent no-tools and 81.2 with tools, matching Sonnet 4.6's baseline. On Humanity's Last Exam with tools, Sonnet 5 scored 57.4 percent, close to Opus 4.8's 57.9. The one number Anthropic keeps quoting is the GDPval-AA v2 knowledge-work index: Sonnet 5 at 1,618, Opus 4.8 at 1,615. That is the first time a Sonnet has crossed an Opus line on any of Anthropic's public knowledge-work benchmarks.

It can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models.— Anthropic, Sonnet 5 launch post

The tokenizer footnote

Simon Willison published his usual same-day measurement of a new Anthropic model on Tuesday afternoon and found the news the launch page did not surface. Sonnet 5 ships with a new tokenizer that produces roughly 30 percent more tokens for identical text than Sonnet 4.6. His measurements: 1.42 times as many tokens for the English text of the Universal Declaration of Human Rights, 1.33 times for Spanish, 1.27 times for a 4,279-line Python file, and 1.01 times for Simplified Mandarin. English workloads pay Sonnet 4.6 economics for equivalent prose even at the $2/$10 promo rate. Willison also reported that Sonnet 5 mistook the pelican in his standard pelican-on-a-bicycle SVG evaluation for a goose. Anthropic also removed the temperature, top_p, and top_k sampling parameters from the API and told developers to use adaptive thinking instead.

The same input text produces approximately 30% more tokens than on Claude Sonnet 4.6. Effectively a 30% price increase.— Simon Willison, simonwillison.net, 2026-06-30

Why the mid-tier is the story

Sonnet 5 is the new default seat for both free-tier and Pro users at claude.ai. The median Claude user, the one who does not think about model selection, now starts on the model that beats Opus 4.8 on GDPval-AA v2 and hits 63.2 on SWE-bench Pro. Anthropic's price move against its own Opus tier is aggressive: Sonnet 5 at $2/$10 promo undercuts Opus 4.8's $5/$25 by a factor of nearly three, and against GPT-5.5 Sol at $5/$30 by two and a half. The Anthropic strategy on display is volume: capture the enterprise renewal negotiation before the customer builds a distillation pipeline against the outgoing tier, which is exactly the workflow The Information described inside Amazon a day earlier.

The safety scores are what let the pricing move. Sonnet 5's system card, per Willison's read, describes the model as significantly less capable at cyber tasks than Mythos 5, which lets Anthropic ship it globally without the June 12 Commerce Department gate that keeps Mythos 5 and Fable 5 offline outside Annex A. The Firefox exploit-development capability eval registered 0.0 percent full-success. Anthropic added that Sonnet 5 was not trained on cybersecurity workflows. That is the concession that made a $2/$10 English-workload discount politically feasible.

A model that knows when to say no is just as important as one that knows how to build.— Fabian Hedin, Lovable co-founder, in the Anthropic testimonials

Launch design partners and the customer story

Anthropic named Zapier senior engineer Daniel Shepard and Lovable co-founder Fabian Hedin among the launch design partners. Shepard's line about the model finishing workflows that used to stall halfway is the exact framing Anthropic wants enterprise buyers to focus on. The company also included a Rust-engineering testimonial from Neel Chotai describing an unprompted end-to-end debug in which Sonnet 5 wrote a reproducing test, implemented the fix, and stashed the branch without hand-holding. The launch materials also emphasized measurable improvements in observed rates of what Anthropic calls undesirable behaviours during agent runs. Anthropic did not publish a full-length system card at launch, so third-party red-teamers will spend the week producing their own numbers on the safety claims.

Claude Code 2.1.197, tagged at 17:56 UTC the same day by Anthropic staff engineer Ashwin Ram, gates access to Sonnet 5 in the terminal on the client-version bump. Claude Code installs below 2.1.197 fall back to Sonnet 4.6 regardless of account entitlements. Pydantic AI 2.2.0 landed four minutes before midnight UTC with Sonnet 5 identifiers merged, becoming the first non-Anthropic agent framework with the model tagged in production.

Big Picture

Tuesday was Anthropic's mid-tier price move against a Chinese trillion-parameter open weight and a Commerce-Department policy wall. Claude Sonnet 5 landed at $2/$10 promo pricing through August 31 with SWE-bench Pro at 63.2 percent, Terminal-Bench 2.1 at 80.4, OSWorld-with-tools at 81.2, and GDPval-AA v2 at 1,618, the first Sonnet score to edge an Opus number on a knowledge-work index. Simon Willison measured the new tokenizer at roughly 1.4x English tokens per unit of prose, cancelling the headline discount for English workloads. The model became the default seat at claude.ai for free and Pro users the same day. Anthropic published Claude Science alongside the model launch, sixty-plus preconfigured research skills, Nvidia BioNeMo integrations across Evo 2, Boltz-2 and OpenFold3, connectors into UniProt, PDB, ChEMBL and GEO, and a $30,000-Claude-credit-plus-$2,000-Modal grant program for approved projects.

Beijing answered inside 24 hours. Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts model with 48 billion active parameters per token and a one-million-token context, released under MIT license on Hugging Face. Meituan claims both pretraining and inference ran end to end on a 50,000-card domestic Chinese ASIC cluster tied downstream to Huawei's Atlas-950 SuperPods and Collective Communication Library. LongCat-2.0 tops Gemini 3.1 Pro and GPT-5.5 on SWE-bench Pro and Terminal-Bench 2.1 while trailing on IFEval. DeepSeek published DSpark the same day, an MIT-licensed speculative-decoding framework claiming per-user speedups of 60 to 85 percent and throughput gains up to 661 percent on DeepSeek-V4-Pro/Flash, Gemma, and Qwen, outperforming EAGLE3 and DFlash. Google filled the fast-cheap corner of its own image and video stack with Nano Banana 2 Lite at $0.034 per 1K image / 4-second generation and Gemini Omni Flash at $0.10 per output-second of video.

Datacenter capital markets ran their heaviest single day of the AI cycle. Digital Realty bought Blackstone's 64 percent blended interest in three fully leased Northern Virginia hyperscale campuses for $3.5 billion, priced against a gross portfolio value of $7.8 billion at a stabilized cap rate above 6.5 percent, then Blackstone priced a 12.3-million-share secondary offering at $185 the same afternoon to convert the stock leg to cash on July 1. Aligned Data Centers priced $1.19 billion of asset-backed securities out of its fifth master trust series, the largest data-center master-trust issuance in market history. Ionic Digital, the Celsius bankruptcy successor, raised $400 million at a $2 billion pre-money valuation ahead of its Nasdaq direct listing under IOND. Etched confirmed a $5 billion post-money valuation and $1 billion of booked contracts for its transformer-only inference silicon, with Karpathy, Hinton, Fei-Fei Li, Druckenmiller and Thiel on the cap table. Alto Infrastructure broke ground on a 70-megawatt AI campus in Granada targeting 250 kW rack densities. Amazon Web Services put down $1 billion for a fully in-house forward-deployed engineer organization under Francessca Vasquez, matching OpenAI's $4B and Anthropic's $1.5B FDE joint ventures without a private-equity partner.

Policy tracked the industrial-policy edges. Samsung and SK Hynix walked back the 800-trillion-won memory-buildout timelines they announced Monday, telling analysts the 400-trillion-won-per-company commitments come without firm timelines and citing 2028 Chinese DDR5 mass-production risk. Taiwan's Keelung District Prosecutors Office raided nine sites including Super Micro Taipei, Chief Telecom and Albatron Technology over an alleged $2.5 billion Nvidia-chip smuggling ring routing servers through Southeast Asian shell companies to China; Super Micro shares fell as much as 9 percent in New York. Colorado's landmark Senate Bill 24-205 reached its official June 30 effective date already repealed and replaced by SB 189, which strips the algorithmic-discrimination duty of care and impact-assessment filing in favour of a narrower disclosure regime effective January 1, 2027. X launched a hosted Model Context Protocol server for Claude, Cursor and Grok Build with read scope only, explicitly blocking autonomous posting. Hugging Face Daily Papers put University of Washington, Leeds and Allen AI's Agentic Abstention at the top with 120 upvotes, followed by Tsinghua's LiveEdit real-time streaming video editor at 70 and Intern Science's Agents-A1 35B MoE at 68.

More Today

26 stories · 2026-06-30
TechPolicy

Meituan open-sources 1.6T LongCat-2.0 trained end-to-end on a 50,000-card domestic Chinese cluster

The Decoder · 2026-06-30 · Source ↗ ·

Meituan's 1.6-trillion-parameter mixture-of-experts model with roughly 48 billion active parameters per token was trained on more than 35 trillion tokens end-to-end on domestically manufactured AI ASICs. It tops Gemini 3.1 Pro and GPT-5.5 on SWE-bench Pro and Terminal-Bench 2.1 while trailing on IFEval.

Meituan, still best known outside China as a food-delivery giant, published LongCat-2.0 on Tuesday and turned the release into a geopolitical statement. The model has 1.6 trillion total parameters and roughly 48 billion active per token, was trained on more than 35 trillion tokens, and ran on a cluster of more than 50,000 domestically manufactured AI ASICs. The South China Morning Post's Hong Kong afternoon file quoted Meituan calling LongCat-2.0 the country's largest artificial intelligence model trained entirely on home-grown hardware. Weights and configs sit on Hugging Face under the meituan-longcat organization under an MIT license.

The hardware detail is what makes the release more than a benchmark drop. Meituan did not formally name the chipmaker, but downstream reports tie the cluster to Huawei's Collective Communication Library and Atlas-950 SuperPods. DeepSeek's V4-Pro used domestic chips only for inference and still leaned on Nvidia for pretraining. LongCat-2.0 is Meituan's claim that Chinese fabs and packaging can now carry an entire trillion-parameter run start to finish. SCMP called it the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing-power cluster.

LongCat-2.0 has demonstrated that we now have the capability to train large-scale models on domestic computing clusters.— Meituan, technical brief, via The Decoder

The benchmark picture

LongCat-2.0 tops Gemini 3.1 Pro and GPT-5.5 on SWE-bench Pro and Terminal-Bench 2.1, both agentic-coding tests that reward long-horizon planning and tool use. It trails Gemini and GPT-5.5 by a wide margin on IFEval, and reasoning-heavy evaluations still favour Opus 4.8 and GPT-5.5 Sol. The mix is familiar: Chinese open-weight coding models keep landing near the frontier on structured code tasks and lagging on instruction following. The router selects a curated set of experts per token, so real activation stays inside what one domestic-chip cluster can serve for inference.

The industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster.— Meituan, quoted by South China Morning Post

Why it matters for the export-control map

Commerce Secretary Howard Lutnick's June 27 letter only partially unwound the June 12 export-control regime that keeps Claude Fable 5 offline globally and Mythos 5 offline outside Annex A. LongCat-2.0 is now the first trillion-parameter-class model that any research team on earth can install and fine-tune under MIT license regardless of jurisdiction. The Chinese Ministry of Industry and Information Technology framed the release in state-media coverage as validation of the June export-control response. Austrian Digital State Secretary Alexander Pröll's June 29 letter to EU Tech Commissioner Henna Virkkunen asking Brussels to pursue the strategic establishment of Anthropic within the EU looks even more urgent this morning, because Beijing has demonstrated an alternative path Europe has not built.

Independent verification is limited by one caveat The Decoder flagged: at publish time the weights on Hugging Face were still being staged, and finance press coverage of the training-compute number relied on Meituan's technical brief. If the 1.6T end-to-end domestic-silicon training claim holds up under community reproduction, the argument that US export controls prevent frontier Chinese scaling gets harder to make.

ApplicationTech

Anthropic launches Claude Science with BioNeMo integrations and a $30,000 credit grant program

Anthropic · 2026-06-30 · Source ↗ ·

Sixty-plus curated research skills, Nvidia BioNeMo integrations covering Evo 2, Boltz-2 and OpenFold3, connectors into UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL and GEO, and a grant program offering up to $30,000 in Claude credits plus $2,000 in Modal compute per approved project. Anthropic said twice in the launch materials that Claude Science is not a new AI model.

Anthropic paired the Sonnet 5 launch with Claude Science, a beta workbench aimed at working researchers rather than developers. The app runs on macOS and Linux, connects to remote machines through SSH or HPC login nodes, and can burst to Modal for on-demand GPU. Sixty-plus preconfigured skills and connectors cover genomics, single-cell analysis, proteomics, structural biology and cheminformatics. Users can spawn specialist agents that report to a generalist coordinating agent, a pattern Anthropic previously used only internally.

The Nvidia partnership is the substantive integration. Claude Science ships with the BioNeMo Agent Toolkit and access to Evo 2, Boltz-2 and OpenFold3. Data-side connectors reach UniProt, the Protein Data Bank, Ensembl, Reactome, ClinVar, ChEMBL and GEO. Anthropic emphasized that large or sensitive datasets stay put on the researcher's HPC login node or workstation. Only the context each analysis step needs travels to Claude, so protected health information and unpublished sequences remain outside the vendor perimeter.

Every output carries an auditable history of how it was made, so you can validate and reproduce the results.— Anthropic, Claude Science launch page

Reproducibility as differentiator

Every figure the workbench generates comes bundled with the code that produced it, the message history that shaped the analysis, and any Modal jobs the pipeline launched. Anthropic is selling peer-review-grade reproducibility rather than chat-window convenience. Rebecca Bellan's TechCrunch report frames the launch as Anthropic betting on workflow orchestration rather than a science-specific model. The company confirmed twice in the launch materials that Claude Science does not ship with a new model and is not a more capable model for biology.

Grants and beta access

Access ties to Claude Pro, Max, Team and Enterprise subscriptions, with academic-institution and nonprofit discounts on Team seats. Anthropic committed to fund 50 Claude Science projects with up to $30,000 in Claude credits each plus $2,000 of Modal compute, targeting postdoctoral and graduate biomedical research. Applications close July 15. Awards land by July 31. Launch users include Allen Institute neuroscientist Jérôme Lecoq and Stephen Francis's group at the UCSF Brain Tumor Center. Anthropic quoted a customer saying an independent research group had validated the workbench's outputs and confirmed it could produce both rapid and robust analyses.

Not a new AI model and not a more capable model for biology.— Anthropic launch materials, repeated twice

Claude Science joins Anthropic's existing government and defence workbench products in the enterprise catalogue. The IPO-track read is that vertical products help defend the Claude API's premium pricing against Chinese open-weight competition, LongCat-2.0 included. Google DeepMind's isomorphic labs, OpenAI's o1-Bio and Meta's ESMFold successors have all pitched biosciences narratives around larger foundation models. Anthropic bet on plumbing.

Tech

DeepSeek open-sources DSpark speculative-decoding for 85 percent per-user speedup and 661 percent throughput

The Decoder · 2026-06-30 · Source ↗ ·

The DSpark drafter drafts multi-word chunks and lets the target model batch-verify them, with adaptive verification depth tuned to fleet load. It outperforms EAGLE3 and DFlash across math, code and chat, works on Gemma and Qwen not just DeepSeek, and ships under MIT license.

DeepSeek released DSpark on Tuesday, an inference-side framework designed to extract more tokens per second from the Chinese hardware still lagging Nvidia H200 and Blackwell on raw memory bandwidth. The framework is MIT-licensed on GitHub and Hugging Face and shipped the same week Beijing kept insisting that inference cost, not compute cost, is the binding constraint under the June 12 export controls. The Decoder's Matthias Bastian tracked both the technical detail and the political frame.

The mechanism: a small drafter proposes token candidates, a larger target model verifies them, and the framework batches those verifications rather than doing them one token at a time. DSpark's twist is that the drafter proposes multi-word chunks rather than single tokens, and the verification depth adjusts on the fly based on the confidence signal and current compute load. When the fleet is quiet, DSpark verifies deeper drafts; when it's under load, shorter drafts keep queue times low.

Most LLMs generate text one word at a time. That leads to low GPU utilization and long wait times for lengthy responses.— DeepSeek, DSpark paper

The numbers

DeepSeek's paper reports per-user response-time gains of 60 to 85 percent and throughput improvements of up to 661 percent. Tests ran on DeepSeek-V4-Flash and DeepSeek-V4-Pro. Beyond DeepSeek's own models, the framework was validated on Google DeepMind's Gemma and Alibaba's Qwen. That cross-family reach matters: the drafter design generalizes across architectures rather than binding to one. DSpark outperformed EAGLE3 and DFlash, the two well-known open speculative-decoding drafters, across math, code and chat benchmarks.

DSpark enables performance tiers that were previously unattainable, shifting the Pareto frontier.— DeepSeek, DSpark paper

The MIT license is aggressive. Anyone can ship DSpark inside a commercial product without strings. DeepSeek released DeepSeek-V4-Pro weights and inference stack together, which puts a frontier open-weight coding model plus a 60 to 85 percent effective throughput gain on Hugging Face for anyone to serve. Layered on top of Meituan's LongCat-2.0, the same-day open-weight aggregate picture in China's ecosystem shifted meaningfully.

Bastian raised the economic point that freed-up compute rarely stays free. In practice, cheaper tokens get absorbed immediately by longer contexts, more agentic loops, or additional users. DSpark is likely to accelerate end-user features that had previously assumed frontier-model quality was available only from Anthropic and OpenAI at scale. The tight coupling to llama.cpp is separate work: b9829-b9835's DFlash v2 sliding-window-per-layer-types has already begun landing in Ollama and the broader inference-runtime stack.

ApplicationEconomic

AWS puts $1 billion into a fully in-house forward-deployed engineer org under Francessca Vasquez

TechCrunch · 2026-06-30 · Source ↗ ·

Amazon Web Services launched a $1 billion internal forward-deployed engineer organization on Tuesday, funded entirely from balance-sheet capital with no PE co-investor. That distinguishes it from OpenAI's $4 billion joint venture and Anthropic's $1.5 billion joint venture, both of which took private-equity partners.

Amazon Web Services launched a $1 billion forward-deployed engineer organization on Tuesday, Russell Brandom at TechCrunch reports, following the same Palantir-invented consulting model that OpenAI and Anthropic have already commercialized. Amazon's twist is structural: the FDE org is fully internal, funded from AWS balance-sheet capital with no outside partner. That differentiates it from OpenAI's $4 billion FDE joint venture with a private-equity co-investor and Anthropic's $1.5 billion joint venture, also PE-backed. Amazon has decided the customer wallet share is worth capturing in-house even at the higher fixed cost.

Francessca Vasquez, AWS's VP of Frontier AI, leads the organization. She told TechCrunch that customers leave AWS FDE deployments with both new solutions and new engineering capabilities. That echoes Palantir's original FDE thesis: temporary engineer embedding, technology reusability and knowledge transfer to the client. It also puts Amazon in a different frame from the pure billable-hour consultancies that Deloitte public-sector lead Jason Manstof described at last week's town hall as a thin sliver of the 2035 revenue picture.

Customers leave AWS FDE deployments with both new solutions and new engineering capabilities.— Francessca Vasquez, AWS VP of Frontier AI

The third major FDE launch in seven weeks

Amazon's move is the third major FDE launch since mid-May, and it lands on a pattern that Bloomberg and The Information have both traced: the moat around frontier-model deployment is the humans who install and tune the systems, not the API itself. Amazon is betting keeping those humans on payroll produces better long-term data flywheels and cross-sell into the AWS enterprise base. The organization will target Fortune 500 accounts and government customers. Vasquez confirmed the initial hire cohort pulls from AWS ProServe, Amazon's existing consulting arm, plus outside recruits from Palantir and Accenture.

The FDE approach was pioneered by Palantir and emphasizes temporary engineer embedding, technology reusability, and knowledge transfer to client organizations.— TechCrunch, summarizing AWS positioning

The $1 billion budget is the largest single AI-services commitment Amazon has publicly disclosed since its Anthropic infrastructure spend. A June McKinsey brief on outcome-based AI contracts noted more than 30 percent of the firm's book is now sold on outcomes rather than time and materials. Customers are pulling for the FDE model despite its labour intensity. Amazon's structural choice not to bring in outside capital signals that the AWS balance sheet reads the customer LTV as strong enough to justify carrying the personnel cost internally.

EconomicTech

Etched confirms $5B valuation with $1B booked, angels Karpathy Hinton Fei-Fei Li Druckenmiller Thiel

TechCrunch · 2026-06-30 · Source ↗ ·

The transformer-only inference ASIC Harvard dropouts sold to Stripes on a December $500M round, disclosed Tuesday, alongside $1 billion of already-booked customer contracts. Jane Street, Hudson River Trading, Two Sigma, VentureTech Alliance and Ribbit are all in.

Etched confirmed on Tuesday that its December 2025 $500 million round valued the company at $5 billion post-money, and disclosed for the first time that the order book contains $1 billion in booked customer contracts for inference systems powered by its chip. Julie Bort's TechCrunch report names Stripes as the lead investor. Total funding to date reaches $800 million. The cap table now includes Jane Street, Hudson River Trading and Two Sigma alongside VentureTech Alliance and Ribbit Capital, plus a slate of individual angels that reads like a deep-learning hall of fame: Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Mistral co-founder Arthur Mensch, Cognition co-founder Scott Wu, Stanley Druckenmiller and Peter Thiel.

Co-founders Gavin Uberti (CEO) and Robert Wachen (president) dropped out of Harvard in 2022 and became Thiel Fellows before launching Etched. Their bet: build a transformer-only ASIC and give up on general-purpose flexibility to unlock inference throughput and cost per token that Nvidia's Hopper and Blackwell architectures cannot match on transformer workloads. The company now sells full frontier inference clusters, combining custom silicon, racks and software into an integrated system rather than shipping bare chips.

Etched has booked $1 billion for the inference systems powered by its chip.— TechCrunch, 2026-06-30

Where this fits in the Nvidia-alternative narrative

The $1 billion contracted-revenue disclosure is the single most important data point for the Nvidia-alternative thesis in 2026. Groq raised $650 million at $7 billion last week. Cerebras IPO'd in May at a $95 billion peak market cap. Rebellions closed a pre-IPO round at $2.3 billion in March. Etched now enters the same category. TechCrunch positions the disclosure inside a broader thesis that AI-inference silicon is emerging as the single most defensible pocket of AI infrastructure economics precisely because customers care more about cost per generated token than benchmark peaks.

Etched specialises in frontier inference clusters, systems combining custom chips, racks, and software designed for faster, cheaper AI model inference.— TechCrunch

Uberti and Wachen's story is also an exception to the general narrative that the Trump administration's export controls are pushing frontier AI toward incumbents. Etched has ridden the same wave that made Micron the memory darling of 2026, and the $800 million war chest positions the company for a Series D or a possible direct listing in early 2027. The angel roster mixes deep-learning academic figures with the largest quant-trading balance sheets on Wall Street. That is not an accident: quant firms buy inference silicon at scale for the same reason model labs do, and Jane Street and Two Sigma writing checks alongside Karpathy is the clearest signal to date that transformer-only silicon has crossed the venture-to-strategic threshold.

TechApplication

Google ships Nano Banana 2 Lite at $0.034 per image and Omni Flash at $0.10 per video-second

The Decoder · 2026-06-30 · Source ↗ ·

Nano Banana 2 Lite (gemini-3.1-flash-lite-image) generates 1K-resolution images in four seconds. Gemini Omni Flash takes text, images and reference clips and outputs up to 10 seconds of video. The recommended pipeline is Lite images fed as references into Omni Flash for animation.

Google DeepMind rolled Nano Banana 2 Lite and Gemini Omni Flash to production on Tuesday. The pair fills in the fast, cheap corner of Google's image-and-video model matrix, sitting under the higher-quality Nano Banana 2 and Veo lines. TechCrunch's Lucas Ropek and The Decoder's Maximilian Schreiner both covered the rollout, and Google separately reclassified the original Nano Banana on the Gemini 3.1 Flash backbone as a legacy model on the same day.

Nano Banana 2 Lite, whose API name is gemini-3.1-flash-lite-image, generates a 1K-resolution image in about four seconds at $0.034 per image. Google positioned it for high-throughput developer pipelines rather than premium content generation. Simon Willison ran it through his raccoon-with-ham-radio Where's-Waldo prompt in AI Studio, called the output better than the earlier Nano Banana models he tested in April, but noted the model spelled "Forest Festival" wrong in two different ways within the same image. Google's own copy calls the model reliable at prompt following, character rendering and readable text in generated images despite the speed focus.

Fastest and cheapest Gemini image model, engineered for velocity and scale.— Google, Nano Banana 2 Lite launch page

Gemini Omni Flash is the bigger jump

Omni Flash takes text, images and reference video clips of up to three seconds, and produces up to 10 seconds of video output at $0.10 per output-second. Google flags conversational video editing in natural language and text-plus-graphics synchronization with video actions as the differentiators. Limitations Google acknowledges include no audio references, no scene extensions and limited character consistency across scene changes. The three-second video-reference input is described as unreliable in the current build.

The recommended workflow: generate a still with Nano Banana 2 Lite, hand it to Omni Flash as a reference and let it animate. Google is packaging the two models as a text-to-motion chain rather than as competing endpoints. Availability spans Google AI Studio, the Gemini API and the Gemini Enterprise Agent Platform. Search AI Mode, the Gemini app, NotebookLM, Google Photos, Stitch, Google Flow and Google Ads rollouts start rolling through the same day.

Generate images with Nano Banana 2 Lite and pass them as references to Gemini Omni Flash, which then animates them into video.— Google, on the recommended pipeline

Pricing puts Google inside Runway and Kling's territory on the video side. At $0.10 per second, a 10-second output costs $1.00. Add eight Lite reference stills and the total lands at $1.27 for a produced text-to-video-with-references sequence. That is fast-food pricing for the segment. For enterprise buyers evaluating video-heavy product-listing generation or marketing-asset workflows, the make-vs-buy math against Runway subscriptions moves.

TechSafety

Hugging Face Daily Papers ranks Agentic Abstention #1 with 120 upvotes, Intern Science 35B Agents-A1 at 68

Hugging Face · 2026-06-30 · Source ↗ ·

University of Washington, Leeds and Allen AI's "Agentic Abstention" benchmark spans 28,000 tasks and shows larger models sometimes abstain worse. Tsinghua's LiveEdit hits 12.66 FPS streaming video edits at state of the art quality. Intern Science's Agents-A1 posts leading scores on SEAL-0, IFBench, HiPhO and FrontierScience-Olympiad from a 35B MoE.

Hugging Face's Daily Papers leaderboard landed on a shape worth noting Tuesday: the top slot is a safety-flavored abstention paper, the number-three slot is an architectural counter to raw parameter scaling, and the number-five slot is threshold-free inference-side compression. The community vote pattern is pushback on the "just scale it" narrative.

At 120 upvotes, Han Luo (Leeds and Southwest Jiaotong University), Bingbing Wen (University of Washington) and Lucy Lu Wang (UW and Allen AI) posted "Agentic Abstention: Do Agents Know When to Stop Instead of Act?" The benchmark is 28,000 tasks across web shopping, terminal environments and question answering. The authors evaluated 13 LLM systems and two agent scaffolds. Headline result: bigger models sometimes get worse at timely abstention, not better. Their scaffold-side intervention, Convolve, needs no parameter updates and raised Llama-3.3-70B's timely-recall rate from 26.7 percent to 57.4 percent.

A reliable agent should recognize that further interaction is unlikely to help and abstain from additional tool calls.— Luo, Wen, Wang, Agentic Abstention abstract

LiveEdit and Agents-A1

Second at 70 upvotes, Tsinghua's LiveEdit ships a real-time diffusion-based streaming video editor. A three-stage distillation pipeline converts a bidirectional foundation into a unidirectional streaming editor and adds an AR-oriented mask cache that reuses per-frame computation. On Wan2.1-T2V-1.3B trained on eight A100s, LiveEdit hits 12.66 FPS at state-of-the-art quality against LucyEdit, InsV2V, VideoCoF, StreamV2V, StreamDiffusion and StreamDiffusionV2. Training set: 20,000 high-quality video-video pairs from Ditto-1M. Inference clocks 7.89 seconds for 81 frames.

Third at 68 upvotes, Intern Science's Agents-A1 is a 35B mixture-of-experts agent trained through a three-stage recipe: full-domain SFT, domain teacher models, and multi-teacher domain-routed on-policy distillation with salient vocabulary alignment. The claimed benchmark deltas are large: GAIA at 96.0 versus Qwen3.5-35B at 59.8; FS-R at 40.0 versus 2.5; IFBench at 80.6 versus 70.2; MLE-Bench-Lite at 43.9 versus 24.2. Compared against 1T-parameter Kimi-K2.6 and DeepSeek-V4-pro, Agents-A1 posts leading scores on SEAL-0 (56.4), IFBench (80.6), HiPhO (46.4), FrontierScience-Olympiad (79.0) and MolBench-Bind (56.8). Weights are open on the InternScience Hugging Face org.

Larger or more capable models sometimes perform worse at timely abstention.— Agentic Abstention abstract

Fourth at 44 upvotes: Meta AI, Duke and Stanford's TUA-Bench, 120 tasks across five terminal-use categories that Claude Code with Opus 4.8 max reasoning tops out at 65.8 percent. Fifth at 43: Tencent WeChat AI and Fudan's ReFreeKV, threshold-free KV-cache compression that adaptively adjusts cache allocation to task complexity without a manual budget knob.

TechEconomic

Digital Realty pays Blackstone $3.5B for three fully leased 96MW Northern Virginia campuses

DatacenterDynamics · 2026-06-30 · Source ↗ ·

$1.2 billion cash plus $2.3 billion of Digital Realty stock for Blackstone's 64 percent blended interest in three fully leased Northern Virginia campuses totaling 288 megawatts. Blackstone priced a $185/share 12.3-million-share secondary the same afternoon to convert the paper to cash.

Digital Realty Trust closed the largest data-center secondary transaction of the AI cycle on Tuesday, buying Blackstone's blended 64 percent equity interest in three Northern Virginia campuses for $3.5 billion. Consideration split into $1.2 billion cash and $2.3 billion of Digital Realty common stock, priced against the June 29 New York Stock Exchange close. Chief Investment Officer Greg Wright led the deal from the Digital Realty side. The portfolio contains two Manassas properties where Blackstone held 80 percent interests and one Sterling property inside the Digital Dulles campus where Blackstone held 50 percent.

Each property carries 96 megawatts of IT capacity for a combined 288 megawatts. All three sites sit fully leased to three separate investment-grade hyperscale customers, and the blended tenant credit rating is AA minus. Lease terms run fifteen years with 3.6 percent annual escalators, above the roughly 3 percent industry norm. The transaction pegs gross portfolio value at $7.8 billion, an expected initial stabilized capitalization rate above 6.5 percent, and reflects the intensity of hyperscaler demand for Loudoun and Prince William County power allocations Virginia stopped accepting new interconnect applications for last quarter.

The demand for digital infrastructure is even stronger today than when we established this joint venture in 2023.— Mike Forman and Greg Blank, Blackstone, joint statement

The secondary offering that closed the loop

The same afternoon, Blackstone priced a 12,310,249-share secondary offering of Digital Realty common stock at $185 per share, Morgan Stanley sole-underwriting on a firm-commitment basis. The offering is scheduled to close July 1. The mechanics let Blackstone convert stock consideration to cash without waiting a customary lock-up window that would apply to unregistered shares issued in an M&A transaction. The $185 pricing represents a 3.6 percent discount to the June 29 close, narrow versus comparable REIT secondaries, reflecting institutional appetite for hyperscale data-center exposure. Digital Realty is not offering shares and receives no proceeds.

We continue to work together across the remaining data center investments in our joint ventures.— Greg Wright, Digital Realty CIO

Digital Realty stock fell 4.5 percent Tuesday on the New York Stock Exchange as the market absorbed the 12.3-million-share dilution. Stifel maintained a $235 twelve-month price target. BMO, JPMorgan and Barclays kept outperform ratings with targets ranging $225 to $245. Digital Realty said the deal is immediately accretive to core funds from operations per share once the assets stabilize in 2027 and 2028, reiterated 2026 core FFO guidance of $8.20 to $8.35 per share, and treated the AA-minus tenant credit rating as investment-grade contracted revenue for balance-sheet purposes. Sources close to the deal said Blackstone realized roughly $850 million of net gains on the two-and-a-half-year hold.

PolicyEconomic

Samsung and SK Hynix quietly walk back Monday's 800 trillion won memory-buildout timelines

Korea Times · 2026-06-30 · Source ↗ ·

Twenty-four hours after President Lee Jae Myung unveiled the joint 800 trillion won ten-year memory package, both chipmakers told analysts the 400 trillion won commitments come without firm timelines. Samsung Vice Chairman Jun Young-hyun: "We have to be cautious about the year of 2028."

Twenty-four hours after Korean President Lee Jae Myung stood alongside Samsung Executive Chairman Lee Jae-yong and SK Group Chairman Chey Tae-won to announce a combined 800 trillion won (roughly $516.4 billion) memory-fabrication and packaging buildout in Gwangju and South Jeolla Province, both chipmakers walked back the ten-year timeline that framed the Monday announcement. The Korea Times reported Tuesday afternoon that Samsung and SK Hynix deliberately omitted specific investment timelines in written follow-up materials, saying the aggregate 400 trillion won commitments per company are provided as guidance and may change depending on market and business conditions.

The framing separates the memory package from prior Samsung capex disclosures. It also gives the two firms cover to slow deployment if the AI memory cycle turns before Gwangju fabs come online. Samsung P4 phase three is targeted for 2027, P4 phase four for 2028, and P5 phase one for 2028. SK Hynix M15X reaches production in 2027 and the two Yongin fabs enter production in 2028. Samsung Vice Chairman Jun Young-hyun and SK Hynix CEO Kwak Noh-jung were named in the follow-up disclosures.

We have to be cautious about the year of 2028.— Jun Young-hyun, Samsung Vice Chairman, in written analyst message

Why 2028 is the caveat

Jun cited caution about 2028, when Chinese memory suppliers are expected to reach mass production of DDR5-grade DRAM potentially at 30 to 40 percent below current spot pricing. Jefferies had used Monday's announcement to lift its DRAM contract-price forecast to plus 40 to 50 percent in Q3 2026 and plus 30 to 40 percent in Q4. If Samsung and SK Hynix defer wafer starts at the new Gwangju sites, the 2027 to 2028 supply picture Jefferies bought into becomes more of a 2029 to 2030 problem, extending the current DRAM up-cycle by roughly a full year.

Long-term investment plans are provided as guidance based on current market conditions and may change depending on future market and business conditions.— Samsung and SK Hynix, joint written statement

Samsung shares closed Tuesday's Seoul session down 1.4 percent and SK Hynix down 2.1 percent as the sell side digested the timeline optionality. The Korean won weakened 0.3 percent against the dollar. Government officials said the flexibility is standard for multi-decade industrial commitments and does not affect the ten trillion won of near-term equipment orders Samsung and SK Hynix committed to under the Monday memorandum. Public-equity analysts covering the two chipmakers read the walk-back as evidence the initial 800 trillion won figure was closer to a political ceiling than a binding commitment. The Lee administration has been under pressure to deliver a memory-sector answer to the Trump administration's June 12 Anthropic export-control regime and the parallel CHIPS Act 2.0 reauthorization now before Congress.

PolicyTech

Taiwan prosecutors raid Super Micro Taipei, Chief Telecom and Albatron over alleged $2.5B Nvidia smuggling

Japan Times / Bloomberg · 2026-06-30 · Source ↗ ·

Keelung District Prosecutors Office simultaneously raided nine sites including Super Micro's Taipei office, colocation operator Chief Telecom and Taipei-listed distributor Albatron Technology, plus six residences, over an alleged shell-company chain routing Nvidia-equipped servers to China. Super Micro shares fell as much as 9 percent.

Taiwan expanded its criminal investigation of Nvidia AI-chip diversion to China on Monday with simultaneous raids on nine sites, including the Taipei office of Super Micro Computer and the offices of Nasdaq-listed Super Micro affiliates Chief Telecom, a Taiwanese data-center operator, and Albatron Technology, a Super Micro distributor listed in Taipei. Six residences were searched. The Keelung District Prosecutors Office confirmed the operation in a public statement, describing it as a widening of an ongoing probe into the alleged illegal export of Super Micro servers containing Nvidia AI accelerator chips.

Prosecutors have alleged the defendants orchestrated a conspiracy to funnel roughly $2.5 billion worth of Nvidia-equipped servers through shell companies in Southeast Asia, ultimately destined for the Chinese market. The Monday action builds on May arrests, when Taiwanese authorities detained three individuals for falsifying export documents involving Super Micro servers equipped with Nvidia AI chips. Investigators believe the suspects successfully shipped at least one batch of Nvidia AI chips to China via Japan and attempted to export approximately 50 servers that were seized before leaving the island. Super Micro said in a Monday evening statement in Taipei that it is cooperating with Taiwanese authorities.

Monday's operation spanned 12 locations, including six private residences and the offices of three companies.— Keelung District Prosecutors Office, statement

The supply chain shape

The three companies raided sit at different points in the Nvidia server supply chain. Super Micro's Taipei office coordinates the company's Taiwanese contract-manufacturing pipeline. Chief Telecom operates colocation facilities in northern Taiwan. Albatron Technology is a memory and graphics-card manufacturer that distributes Super Micro chassis to Taiwanese enterprise customers. Prosecutors have not publicly named the six individuals whose residences were searched. Bloomberg sources told the wire that the six are believed to be senior sales executives, freight forwarders and customs brokers who allegedly coordinated shipment routing. Taiwanese customs authorities have said the alleged shell-company chain sent servers to intermediaries in Malaysia and Vietnam that then transshipped to Hong Kong or the mainland.

A conspiracy to funnel roughly $2.5 billion worth of Nvidia-equipped servers through shell companies in Southeast Asia.— Prosecutors' filing, per Bloomberg

Super Micro shares fell as much as 9 percent in New York trading before recovering some of the losses to close 6.4 percent lower. The Trump administration has been pressuring Taipei to tighten enforcement, and the June 12 US export-control expansion around Anthropic Mythos 5 explicitly framed Taiwanese distributor networks as a strategic supply-chain risk. The Monday raids indicate that pressure is producing observable enforcement outcomes. Super Micro's compliance disclosures had not previously mentioned an active Taiwanese criminal investigation of the company. Wedbush analyst Rob Cihra cut his 2027 revenue forecast for Super Micro by three percent to reflect potential shipment restrictions on the company's Taiwanese contract-manufacturing capacity. Taiwan is also considering a legislative package that would make unauthorized export of AI chips a criminal offense under Taiwanese law.

Policy

Colorado's landmark AI Act reaches its June 30 effective date already repealed and replaced

Greenberg Traurig / National Law Review · 2026-06-30 · Source ↗ ·

Governor Jared Polis signed SB 189 on May 14, repealing and replacing SB 24-205 mere weeks before its effective date. The replacement drops the duty of care, risk-management programs and impact assessments in favour of a disclosure-only framework taking effect January 1, 2027.

Colorado Senate Bill 24-205, the Colorado AI Act, reached its official effective date Tuesday, and the law never took effect. Governor Jared Polis signed SB 189 on May 14 repealing and replacing the entire statute with a narrower regime that will now take effect January 1, 2027. The June 30 date matters not because anything happened at midnight, but because Colorado's climb-down closes a two-year experiment in state-level frontier-AI regulation and marks a turning point for similar bills advancing in California, New York, New Jersey, Texas and Virginia.

SB 24-205 was the first US statute to impose a duty of reasonable care on developers and deployers of high-risk AI systems to prevent algorithmic discrimination in consequential decisions covering employment, housing, lending, insurance, healthcare and education. It required annual impact assessments filed with the state attorney general, risk-management programs modelled on NIST's AI RMF, and mandatory consumer disclosures. Penalties reached $20,000 per violation. Governor Polis signed the original law in May 2024 while simultaneously asking the legislature to fix its problems before it took effect.

SB 189 repealed and replaced the Colorado AI Act mere weeks before the original statute's June 30, 2026 effective date.— National Law Review

What survives, what goes

The replacement law limits itself to disclosure requirements around automated decision-making technology that materially influences consequential decisions and drops the risk-management duty, the impact-assessment filing and the algorithmic-discrimination duty of care. Colorado Attorney General Phil Weiser must promulgate implementing regulations by January 1, 2027. Colorado businesses and, notably, the Colorado Chamber of Commerce joined national industry groups arguing the compliance load was uneconomic and the algorithmic-discrimination duty unworkable. The legislature responded twice: SB 25B-004 signed August 28, 2025 pushed the effective date from February 1 to June 30, 2026, and SB 189 replaced the statute entirely.

The new law removes many of the hallmarks of the Colorado AI Act, such as a duty of care, risk management programs, and impact assessments, in favor of a disclosure-based framework.— Davis Wright Tremaine analysis

The Colorado AI Act was the only state law specifically named in President Trump's December 2025 Executive Order as an example of a state law entailing excessive state regulation, and the White House claimed the algorithmic-discrimination duty would force AI models to produce false results. Trump's June 2 executive order on AI innovation and security codified a broader preference for federal, industry-facilitated safety review. The Colorado repeal is now the first material state-level surrender to that framework. California's SB 1047 successors, New York's AI Bill of Rights Act, New Jersey's algorithmic-decisioning bill and Virginia's HB 2094 all watched Colorado's roll-back closely.

PolicyApplication

X ships a read-only MCP server for Claude, Cursor and Grok Build, no autonomous posting

TechCrunch · 2026-06-30 · Source ↗ ·

User-scoped access lets any MCP client search posts, read conversations, look up users, analyse trends and pull real-time data. X explicitly disabled Write API endpoints, a deliberate choice against the industry pattern of write-scope MCP integrations.

X launched a hosted Model Context Protocol server on Tuesday, giving Claude, Cursor, Grok Build and any MCP-compatible client user-scoped access to search posts, read conversations, look up users, analyse trends and pull real-time data using the developer's own X account permissions. The read-only scope is the policy story. X clarified to TechCrunch that the MCP tool is not compatible with X's Write API endpoints, which prevents autonomous posting by an AI agent operating on a user's behalf.

That is a deliberate choice against the industry trend. GitHub, Slack, Notion, Stripe and Salesforce all offer MCP servers with write access, and Claude Code has demonstrated write-scope MCP integrations across Google Workspace and Microsoft Graph. X's decision to withhold write access is partly a policy response to the deepfake, non-consensual synthetic-media and election-integrity concerns that have dominated 2026 platform governance. The FEC remains partisan-deadlocked on AI-in-campaigns rules, 47 states have passed some form of synthetic-media legislation, and the June X platform update explicitly required provenance labels and generation metadata on AI-produced content. Allowing autonomous agents to post on behalf of any account would undermine the labelling regime overnight.

The MCP tool is not compatible with X's Write API endpoints.— X, statement to TechCrunch

Pricing was probably a secondary consideration

X also raised its Write-API pricing recently, with publishing posts costing $0.015 apiece and posting links costing $0.20 each. That suggests economics figured into the decision but was not the primary driver. The launch has significant policy ripple effects. Claude Code users can now pull real-time X context into agent workflows including trend research, competitive analysis and social-listening automation, without X owing agents any writing capability. That is the position that Bluesky and Mastodon have occupied by default.

MCP is an open standard defining how AI models connect to external tools and services.— TechCrunch, on the protocol context

For EU AI Act General-Purpose AI Code of Practice signatories, X's read-only MCP model provides a working template for how platforms can offer AI integration while excluding the highest-risk misuse pathway. xAI, which signed only the Safety and Security chapter of the EU code, is likely to be the first beneficiary of the read-only precedent. The launch positions X as a source of real-time signal for agents without turning it into an autonomous-posting distribution channel, which is exactly the shape most enterprise buyers wanted to see before wiring the platform into agent workflows.

EconomicTech

Aligned prices $1.19B asset-backed securitization backed by Ashburn, Plano and Northlake sites

DatacenterDynamics · 2026-06-30 · Source ↗ ·

Fifth master-trust series since 2023, collateralized by four US Aligned campuses in Ashburn, Plano and Northlake. Morgan Stanley led as sole structuring advisor. Deal closes July 27. Aggregate trust appraised value falls $1.09B to $3.16B as three sites exit and four remain.

Aligned Data Centers priced $1.19 billion of asset-backed securities on Tuesday, the fifth series issued out of the company's master-trust vehicle since the platform launched in 2023. Real estate and tenant lease payments across four US Aligned campuses back the notes: Ashburn, Virginia; Plano, Texas; and two facilities at Northlake, Illinois. S&P Global Ratings' pre-sale analysis described the trust as a special-purpose vehicle bundling revolving debt across multiple bond series drawn from a dynamic pool of hyperscale colocation assets.

Under this issuance, three data centers exit the previous series 2023-2 pool, four data centers stay in the new pool, and aggregate appraised trust value falls by $1.09 billion to $3.16 billion. The four remaining collateral properties saw their appraised value rise $887 million on the back of AI leasing demand and 22 megawatts of new leased capacity. Net leased capacity in the trust decreased by 101 megawatts because of the departing sites. The transaction contains four tranches of Class A1, A2 and Class B notes with an anticipated maturity of five years and a legal final maturity of thirty years. At close, the deal will benefit from a $36.9 million reserve account funded from series 2026-1-L/C proceeds. Cash-trap and Class A amortization triggers are set at three-month average Class A DSCRs of 1.35x and 1.20x respectively.

The transaction benefits from a diversified pool of hyperscale colocation assets in top US data center markets.— S&P Global Ratings, pre-sale report

The BlackRock consortium context

Aligned's parent BlackRock-led consortium acquired the company in a $40 billion transaction that closed in the first quarter of 2026. The consortium includes Microsoft, Nvidia, xAI and MGX. The June 30 issuance is the second master-trust series priced since the take-private closed, and lifts total outstanding master-trust debt to a level that Aligned executives previously said would target investment-grade ratings triple-B-minus or higher on senior tranches. The master-trust structure lets Aligned refinance and re-lever the same pool of stabilized colocation assets as leases roll and net operating income compounds.

Cash trap trigger and a class A amortization trigger based on the three-month average class A debt service coverage ratio levels of 1.35x and 1.20x.— S&P Global Ratings pre-sale report

Morgan Stanley acted as sole structuring advisor. The deal is slated to close July 27. Real-estate finance groups covering the transaction described the master trust as the emerging financing template for privately owned hyperscale operators shut out of the public REIT capital markets. Rating agencies have leaned on the AA-minus blended credit rating of Aligned's hyperscale tenants and the 15-year weighted-average remaining lease term as the twin anchors of the securitization's investment-grade rating. Aligned currently operates 47 data centers in the Americas and EMEA, with a publicly disclosed project pipeline of over 5 gigawatts of new capacity through 2028.

TechEconomic

Alto Infrastructure breaks ground on 70MW Granada AI campus targeting €3B, 250kW rack density

DatacenterDynamics · 2026-06-30 · Source ↗ ·

Spanish operator formerly Sierra DC broke ground Tuesday on SP01 in Escúzar, Granada. Direct-to-chip liquid cooling, 100MW electrical / 70MW IT at full build-out in late 2029. Total potential investment €3B with roughly €700M on physical infrastructure and the rest through tenant compute capex.

Spanish operator Alto Infrastructure, until recently trading as Sierra DC, broke ground Tuesday on SP01, a purpose-built AI, high-performance computing and cloud-services campus in the Industrial, Technological and Innovation Area of Escúzar, in the province of Granada, southern Spain. The company put the total potential investment at over €3 billion (roughly $3.43 billion). Roughly €700 million (about $800 million) funds physical infrastructure. The balance is expected to flow through tenant capital expenditure on servers, GPUs and advanced compute platforms once the site opens.

The first phase reaches operational status in the summer of 2027 with 10MW of IT capacity, expands to 25MW before the end of 2027, and reaches full build-out at 100MW electrical / 70MW IT in late 2029. Alto Infrastructure will use direct-to-chip liquid cooling and has designed the campus to support rack densities up to 250 kilowatts, an order of magnitude beyond air-cooled colocation. CEO Gísli Kr. described the site as a response to the physics shift of AI training and inference workloads. The Granada region combines available land, upstream grid capacity for fresh renewable and gas-fired generation, and proximity to IFMIF DONES, one of Europe's main fusion-energy research projects. The University of Granada anchors a local technology talent pool.

The most advanced models demand ever-increasing processing capacity and power densities far higher than traditional data centers.— Gísli Kr., CEO of Alto Infrastructure

Iberia versus FLAP-D

Alto Infrastructure joins a wave of Iberian datacenter announcements that positions southern Spain as a rising alternative to the FLAP-D markets of Frankfurt, London, Amsterdam, Paris and Dublin, where new interconnect capacity is now measured in years of queue. Grid connectivity through Red Eléctrica de España and the Andalusian regional operator supports direct transmission-network connection and site-provided generation. Spain's national industrial strategy has quietly reshaped incentives for large data-center anchors, matching German Bundesnetzagentur allocations for the first time in 2026.

The province combines availability of land, access to energy, connectivity and a growing technological ecosystem.— Alto Infrastructure, on Granada

The rebrand from Sierra DC to Alto Infrastructure marks the company's pivot from a traditional colocation identity toward positioning itself as an operator of vertically integrated AI factories, in the vocabulary Nvidia uses to describe the packaged compute, cooling and networking systems it deploys with hyperscale customers. Alto has not disclosed anchor-tenant commitments for the Granada site. The 250 kilowatts per rack design target sits near the top of the current European rack-density curve, matching the requirements of Nvidia's forthcoming Vera Rubin racks and the OpenAI-Broadcom Jalapeño platform expected to deploy at the end of 2026.

EconomicTech

Ionic Digital locks in $400M pre-IPO at $2B, AI colocation revenue overtakes bitcoin mining

DatacenterDynamics · 2026-06-30 · Source ↗ ·

The Celsius bankruptcy successor priced a $400M private placement at a $2B pre-money valuation ahead of a Nasdaq direct listing under IOND. Attestor, Oaktree and Sachem Head led. Q1 2026 saw digital-infrastructure revenue of $44M outrank bitcoin mining's $7.4M for the first time.

Ionic Digital completed a $400 million equity private placement Tuesday, pricing at a $2 billion pre-money valuation ahead of a Nasdaq direct listing under the ticker IOND. New investors Attestor, Oaktree Capital Management and Sachem Head Capital Management led. Citadel and Weiss Asset Management also participated. J.P. Morgan and BTIG served as placement agents. White & Case acted as legal counsel. CEO Andy Stewart said the financing strengthens the capital base and supports continued development of the digital-infrastructure platform.

Ionic runs approximately 120,000 mining rigs across Texas at Garden City (12MW), Rebel (50MW), Stiles (20MW) and East Stiles (30MW). The company plans to expand its Barstow site to 700MW of critical IT capacity by end of 2027. In Q1 2026, digital-infrastructure leasing revenue reached $44.0 million against $7.4 million from bitcoin mining, marking the crossover point at which the AI colocation business overtook the legacy crypto operation as the primary revenue engine. Total Q1 revenue reached $51.4 million. Ionic Digital is the entity that emerged out of the Celsius Network bankruptcy, with the miner distributed to Celsius creditors as part of the reorganization plan in early 2024.

This financing strengthens Ionic Digital's capital base and supports the continued development of our digital infrastructure platform.— Andy Stewart, Ionic Digital CEO

Direct listing, not traditional IPO

The company filed its Form S-1 with the SEC on Monday June 29 seeking to list class-A common stock on the Nasdaq Global Select Market. The move is structured as a direct listing rather than a traditional IPO, meaning the listing itself will not raise new cash for Ionic. The $400 million private placement replaces the fresh capital a conventional IPO underwriting would have generated. The $2 billion valuation values Ionic at roughly 40 times trailing quarterly revenue and roughly 10 times annualized run-rate revenue, in line with the multiples Applied Digital, TeraWulf and Iris Energy carry as they execute the same crypto-to-AI pivot.

Digital infrastructure leasing accounted for $44.0 million, compared with just $7.4 million from bitcoin mining.— Ionic Digital, S-1 disclosure

The Barstow expansion to 700MW by end of 2027 would position Ionic as one of the largest single-site AI colocation providers in Texas outside the hyperscaler self-build footprint. Ionic has publicly committed to selling roughly 60 percent of the Barstow expansion to a single anchor customer under a fifteen-year lease, though management declined to name the anchor in the S-1. The private-placement pool includes Citadel Advisors as a repeat investor from the Celsius emergence trade in 2024. Weiss Asset Management, run by Andrew Weiss, was a distressed-debt participant in the original Celsius bankruptcy. J.P. Morgan will act as designated market maker on the direct listing, expected to price in the second half of July.

Economic

Ex-DeepMind poker AI trio's EquiLibre hits $500M Series A trading S&P 500 for Tower Research

TechCrunch · 2026-06-30 · Source ↗ ·

Prague-based EquiLibre Technologies, founded by three ex-DeepMind researchers behind the landmark DeepStack and Cicero superhuman poker systems, closed a Series A valuing the company at $500M led by Creandum. Already trading S&P 500 and Nasdaq for Tower Research Capital.

EquiLibre Technologies, founded by three former DeepMind researchers behind the landmark superhuman poker AI, closed a Series A valuing the company at $500 million on Tuesday. Creandum led, describing the cheque as its largest single investment on record. Round size was not disclosed. The Series A follows a $10 million seed at a $140 million valuation led by Blossom Capital plus pre-seed backing from Credo. Valuation has tripled and change in roughly a year.

The three founders are Martin Schmid (CEO), Rudolf Kadlec (CTO) and Matej Moravcik (chief scientist), all originally part of the DeepMind Prague team that built DeepStack and Cicero, the poker systems widely considered the AlphaGo of imperfect-information games. EquiLibre incorporated in 2022 and pivoted from a general reinforcement-learning research bet to the specific problem of automated trading in liquid public markets. Schmid told TechCrunch the appeal was simplicity of scoring: the nice thing about trading, he said, is that the scoring is super simple, how much money did the agent make.

The nice thing about trading and markets is that the scoring is super simple: how much money did the agent make?— Martin Schmid, EquiLibre CEO

The Tower Research relationship

The company began with crypto strategies in 2025 and now trades the S&P 500 and Nasdaq for Tower Research Capital, one of the world's largest high-frequency trading firms. That customer disclosure is central to Creandum's underwriting. Cameron Sellers, Creandum VP, told TechCrunch that the potential total addressable market of trading in financial markets is one of the biggest on earth. Schmid pushed back against the pure-money framing, telling TechCrunch his motivation is building new things, not making markets more efficient.

I'm not doing this because I'm excited about making markets efficient. I'm doing this because we are all excited about building new things.— Martin Schmid, EquiLibre CEO

The Series A joins a small but growing cohort of AI-native quantitative trading startups. Rival Qube Research and Technologies raised $1 billion at $8 billion earlier in 2026, and Cognition's coding agents are being repurposed inside multiple hedge funds. EquiLibre's edge, per the pitch, is that Kadlec and Moravcik's counterfactual-regret-minimisation research applies directly to portfolio construction under adversarial market conditions. The Tower Research relationship also acts as a distribution channel: EquiLibre gets execution infrastructure it does not have to build, and Tower gets exposure to a research shop with no legacy portfolio risk.

EconomicApplication

OKX opens agent-hires-agent marketplace with stablecoin settlement and 50 closed-beta providers

TechCrunch · 2026-06-30 · Source ↗ ·

Autonomous agents can hire each other, settle in stablecoins and build on-chain reputations. Compatible clients include Claude Code, Codex, Hermes and OpenClaw. Founder Star Xu: one-person, million-dollar companies will define the coming decade.

Crypto exchange OKX opened its AI-agent marketplace to developers on Tuesday, giving autonomous agents a way to hire each other and settle payments through stablecoins while building on-chain reputations. Jagmeet Singh at TechCrunch reports the launch follows a closed beta with 50 early service providers, including CertiK for security assessments, CoinAnk for pay-per-query market data and GenLayer for dispute-resolution infrastructure. Compatible clients at launch include Claude Code, Codex, Hermes and OpenClaw. An agent running inside any of those harnesses can dispatch a paid task to a marketplace agent, receive the completed output and record the exchange to chain.

OKX founder and CEO Star Xu framed the marketplace with a founder pitch that echoed the venture-capital consensus of 2026: the coming decade will be defined by one-person companies that generate over a million dollars in annual revenue, because every individual effectively gains an unlimited workforce. Chief marketing officer and global managing partner Haider Rafique sized the market as trillion-dollar over the next five years, driven by micropayments and autonomous software.

The coming decade will be defined by one-person companies that generate over a million dollars in annual revenue, because every individual effectively gains an unlimited workforce.— Star Xu, OKX founder and CEO

The ICE backing and India angle

The platform is accessible through OKX's Onchain OS toolkit, which sits inside the existing OKX 150-million-user base. Intercontinental Exchange, owner of the New York Stock Exchange, invested about $200 million in OKX in March at a $25 billion valuation, and the marketplace launch is the first product OKX has shipped since that raise. India is on the strategic-priority list for developer recruitment because OKX suspended its Indian crypto trading services in 2024 and the AI marketplace lets it re-engage the region without touching the still-fraught regulatory environment around crypto trading.

Agentic commerce could become a trillion-dollar market over the next five years, driven by micropayments and autonomous software.— Haider Rafique, OKX chief marketing officer

The launch competes with AgentPay from Coinbase, previously announced at Consensus 2026, and with Base's agent-payment SDK. OKX's edge, per Singh's reporting, is the existing exchange's balance sheet, its Cayman-domiciled compliance stack that avoids US securities complications, and the closed-beta traction that produced quality service providers before public launch. Micropayment rails, until now the bottleneck for agentic commerce, are running on stablecoin infrastructure the exchange already operates. The June 30 cluster of agent-to-agent flows also connects to Mozilla 0DIN's Sunday disclosure of the Claude Code DNS-tunneled supply-chain attack: any marketplace that lets agents pay each other for work opens fresh attack surface on the recipient side.

EconomicApplication

Acti raises $5.3M from BITKRAFT to put agentic AI inside iOS and Android keyboards

TechCrunch · 2026-06-30 · Source ↗ ·

Singapore's Acti launched a Gemini-powered agentic keyboard on iOS and Android and disclosed a $5.3M BITKRAFT-led seed. Users describe cross-app Skills in natural language. 1,000 public Skills were created during early access. Local-first architecture keeps personal context on device.

Singapore-based Acti launched a Gemini-powered agentic keyboard on iOS and Android on Tuesday and disclosed a $5.3 million seed round led by BITKRAFT Ventures. Sarah Perez at TechCrunch reports founder and CEO Young Wang built the product around a specific bet: today's AI agents stay limited because user context stays fragmented across separate apps. The keyboard sits at the input layer every app must share, so an agent embedded there can carry context between mail, messaging and social apps without any per-app integration. Users describe Skills in natural language and Acti generates them on device.

Wang told TechCrunch the local-first architecture keeps personal context on device by default, an important product differentiator against server-based competitors such as Rabbit and Humane. Built-in shortcuts include a T Skill for translation and a C Skill that turns any typed reference to a meeting into a calendar link with attendees. During early access, users created more than 1,000 public Skills that any other Acti user can install from the Skills marketplace.

Today's AI agents are fundamentally limited because user context stays fragmented across separate apps.— Young Wang, Acti founder and CEO

The founding team

CTO Mike Sun was the founding technical lead at Baidu's Yike Album consumer product. CSO Junbo Yang joined from HashKey Capital, the Hong Kong-listed digital-asset firm. Jonathan Huang, general partner at BITKRAFT Ventures, told TechCrunch that BITKRAFT backed Acti because this team has a real shot at owning the next phase of human-computer interaction. BITKRAFT's thesis, forged in gaming and consumer products, is that interaction paradigms shift roughly once a decade, and the keyboard is a plausible next winner in a world where users no longer speak to apps but delegate to agents.

We backed Acti because this team has a real shot at owning the next phase of human-computer interaction.— Jonathan Huang, BITKRAFT Ventures general partner

Acti positions itself against Google's Gboard and Apple's system keyboard, which have both added AI features but not full agentic tool use. The $5.3 million cheque is small by 2026 AI standards but places Acti in the Series A hunting ground for consumer AI. TechCrunch notes the launch coincides with a broader consumer-agent wave that includes Rewind's Pocket recorder ($11 million from Accel and ElevenLabs founder Mati Staniszewski) and Meta's rumoured agentic Ray-Ban Explorer.

SafetyApplication

Proton Lumo 2.0 ships zero-access-encryption chatbot with 76 percent faster response and image gen

TechCrunch · 2026-06-30 · Source ↗ ·

Re-engineered from the ground up with image recognition, image generation, thinking mode and user-controlled persistent memory scoped to Projects. Founder Andy Yen: users no longer need to choose between powerful AI capabilities and meaningful privacy protections.

Proton released Lumo 2.0 on Tuesday, a privacy-first AI chatbot upgrade that Lucas Ropek at TechCrunch reports has been re-engineered from the ground up. Founder and CEO Andy Yen framed the release as a direct rebuttal to the assumption that users must trade privacy for capability. Lumo 2.0 ships with image recognition, image generation, a thinking mode for complex queries, user-controlled persistent memory scoped to Projects, and 76 percent faster response times than the previous version.

Response times matter here because chatbot latency is the single largest driver of user complaints against privacy-first alternatives to ChatGPT and Claude. Proton's architectural bet is zero-access encryption for chat contents and no server-side session logging. Neither Proton nor a hypothetical government subpoena can retroactively access what a user typed. The persistent-memory feature required careful engineering to preserve that guarantee, because the industry pattern is that memory features leak metadata onto the vendor's servers.

Lumo 2.0 demonstrates that users no longer need to choose between powerful AI capabilities and meaningful privacy protections.— Andy Yen, Proton founder and CEO

The policy climate that made this launch land

Yen made the political framing explicit. The release lands during a period when AI privacy has become an enforcement lever. The Rhode Island chatbot-therapy ban signed on June 22, the FTC's active September 2025 formal inquiry into AI child safety at Meta, OpenAI and Google, and the ongoing revelations from the Wired investigation into Meta contractors posing as minors to test rival chatbots have all pushed AI privacy into the mainstream policy conversation.

Lumo 2.0 has been re-engineered from the ground up and the introduction of thinking mode gives it powerful new capabilities.— Andy Yen, Proton

Proton positions Lumo as the safe choice for European users who cannot rely on OpenAI, Anthropic or Google to hold data outside the reach of the CLOUD Act. That positioning gained additional weight after Austrian State Secretary Alexander Pröll's June 29 letter to Henna Virkkunen calling for the strategic establishment of Anthropic within the European Union. Lumo's architecture provides an existence proof that private AI is buildable now, without waiting for European frontier models to arrive. The image-generation feature ships against a market where every hyperscaler is racing to lower cost per generation, so Lumo's differentiation is on encryption, not price. Google's Nano Banana 2 Lite at $0.034 per image landed the same day.

SafetyApplication

OpenClaw open-source agent lands on iOS and Android under the MoltBook shadow

TechCrunch · 2026-06-30 · Source ↗ ·

Peter Steinberger's free open-source agent shipped mobile clients Tuesday. Users pair phones with OpenClaw Gateway to run agents from their pocket. The MoltBook affair, in which researchers exposed viral demos as partially human-driven, still shadows the release.

OpenClaw, the free open-source AI agent program created by Peter Steinberger (now at OpenAI), landed on both iOS and Android on Tuesday, Lucas Ropek at TechCrunch reports. The mobile launch means users pair their phone with OpenClaw Gateway, a routing layer that connects incoming user requests to AI agents and their associated tools, and can then run OpenClaw agents from their pocket. The safety story is the context surrounding the launch.

OpenClaw went viral earlier in 2026 around what became known as the MoltBook incident, a stunt-marketing event in which researchers subsequently exposed that many of the viral agent demonstrations had been partially human-driven rather than autonomously executed by the agent. That controversy produced a wave of industry writing about human-in-the-loop deception in agent demos, and about the trust cost when autonomous-agent demos are quietly staffed by humans behind the curtain. Steinberger's launch of the mobile client comes into that reputational headwind.

The free open source agentic program is finally invading your phone.— TechCrunch, on the mobile launch

The wider agent-safety day

TechCrunch notes users have applied OpenClaw to workflows including coding and meal planning, with some reporting suboptimal results, a framing that echoes the broader 2026 discourse around agent reliability. The mobile release arrives on the same day OKX's AI-agent marketplace opened to developers and confirmed OpenClaw as a compatible client, which means a mobile OpenClaw user could theoretically dispatch paid work to a marketplace agent while at a bus stop. That agent-to-agent flow is precisely the pattern that Mozilla 0DIN's Sunday DNS-tunneled supply-chain attack against Claude Code exposed, in which malicious payloads never enter the visible GitHub repo and instead route through domain lookups the agent trusts.

The MoltBook demonstration was later revealed by researchers as partially human-driven marketing theater.— TechCrunch, on the reputation context

OpenClaw's open-source nature makes it a natural target for the same class of attack. The June 30 launch will accelerate red-team scrutiny even as it accelerates adoption. Steinberger's post on X announced the mobile client. The free-tier open-source posture means the community, not a corporate security team, will be the first line of defence against the next MoltBook or the next Cordyceps. Peter Steinberger has been at OpenAI since March, when he was recruited alongside a cohort of open-source developer-tool authors. He continues to maintain OpenClaw as a personal open-source project.

TechApplication

Claude Code 2.1.197 promotes Sonnet 5 to default with native 1M-token context

GitHub · 2026-06-30 · Source ↗ ·

Ashwin Ram tagged 2.1.197 at 17:56 UTC and flipped the default from Sonnet 4.6 to Sonnet 5. Installs below 2.1.197 fall back to 4.6 regardless of account entitlements. Removes support for temperature, top_p and top_k in favour of adaptive thinking.

Anthropic staff engineer Ashwin Ram tagged Claude Code v2.1.197 at 17:56 UTC on Tuesday, one day after 2.1.196 shipped the streaming idle watchdog. The commit flips the default model in Claude Code from the outgoing Sonnet 4.6 to Claude Sonnet 5, the version Anthropic released to the public at 10 a.m. Pacific with matching news on anthropic.com/news/claude-sonnet-5. Ram gates access on the client version. Every Claude Code install below 2.1.197 falls back to Sonnet 4.6 regardless of whether the account has Sonnet 5 entitlements provisioned, so terminal users who run auto-update opt in to the new model without touching a config file.

The changelog is one paragraph: introducing Claude Sonnet 5, now the default model in Claude Code, with a native 1M-token context window and promotional pricing of $2/$10 per Mtok through August 31; update to version 2.1.197 for access. The engineering delta from 2.1.196 is a config bump that updates the default model identifier and adjusts the context-window ceiling logic that had been capping requests at 200K tokens for Sonnet 4.6 users.

Introducing Claude Sonnet 5: now the default model in Claude Code, with a native 1M-token context window and promotional pricing of $2/$10 per Mtok through August 31.— Claude Code 2.1.197 release notes

What the release removes

The release also removes support for the temperature, top_p, and top_k sampling parameters from the model, replacing them with an adaptive-thinking control that Anthropic turns on by default. Claude Code 2.1.197 exposes no CLI flag to disable adaptive thinking, sending the responsibility for cost controls to per-organization allowance limits. The Anthropic system card notes Sonnet 5 is significantly less capable at cyber tasks than Mythos 5, letting it ship without the export-control gate the Commerce Department imposed on Mythos 5 and Fable 5 on June 12.

The same input text produces approximately 30% more tokens than on Claude Sonnet 4.6.— Simon Willison, on the new tokenizer

Simon Willison flagged the same day that Sonnet 5's new tokenizer produces about 30 percent more tokens on the same input than Sonnet 4.6. His measurements landed on 1.42x for English prose, 1.33x for Spanish, 1.27x for Python, and 1.01x for Simplified Mandarin. Willison read the tokenizer as a hidden price bump on top of the nominal cut. The community-reactions counter on the release shows 17 thumbs up, six party horns, four hearts, and four rockets in the two hours after the tag. The 5-minute streaming idle watchdog from 2.1.196 continues to run, and the org-default model override now targets Sonnet 5 unless overridden.

Tech

ggml-org tags seven llama.cpp builds Tuesday including Asahi Vulkan SPIR-V edit and Gemma E4B fix

GitHub · 2026-06-30 · Source ↗ ·

Jeff Bolz merged xingjianll's PR #24663 into b9846 with a runtime SPIR-V bytecode edit that rolls the matmul bk loop on Apple silicon under Asahi Linux for 22.4 percent faster prompt processing. Johannes Gaessler's b9847 fixes Gemma E4B MTP FlashAttention. jschmied's b9850 registers t_layer_inp for qwen3next unblocking DFlash.

ggml-org tagged seven llama.cpp builds between 08:56 and 19:35 UTC on Tuesday, an unusually dense day even by llama.cpp standards. Jeff Bolz merged PR #24663 at 11:03 UTC as b9846, closing a real Asahi Linux problem: on Apple M-series silicon the fully unrolled fused-multiply-add loop inside the matrix-multiplication shader spilled instructions out of the 12 KB M-series instruction cache. Contributor xingjianll's first draft solved the problem by generating multiple SPIR-V shader variants at compile time. Bolz pushed back on the binary-size cost: 217 megabytes of shader source and 89 megabytes of shipped DLL before, versus 384 megabytes and 135 megabytes after. xingjianll rewrote the fix to modify SPIR-V bytecode directly during pipeline creation at runtime, stripping the loop-unroll hint from the pre-compiled shader only when the driver reports an Asahi-family device. Performance on the M1 Ultra with 48 GPU cores: 22.4 percent gain on prompt processing at 512 tokens, 2 percent regression on 128-token generation. Bolz merged after two review rounds.

vulkan: roll bk loop in matmul for asahi linux · Edit SPIRV directly for Asahi roll bk loop implementation.— b9846 release notes

Gemma 4 E4B and qwen3next

Johannes Gaessler self-merged PR #25148 at 13:16 UTC as b9847, restoring Gemma 4 E4B FlashAttention on CUDA. His prior template-compilation optimization had disabled specialization for GQA ratios 1 and 2, and a subsequent parallelization PR then silently dropped the GQA-ratio-2 specialization entirely from the compiled kernel list. Gemma 4 E4B MTP depends on GQA ratio 2. Gaessler's b9847 adds the specialization back and removes a genuinely unused template declaration to keep binary size neutral. His commit note reads plainly: he was not aware that the template specialization with a GQA ratio of 2 is actually in use so that PR inadvertently broke Gemma 4 E4B MTP.

CUDA: fix Gemma E4B MTP FlashAttention · remove unused template declaration.— b9847 release notes

CISC and ruixiang63 approved and merged jschmied's PR #25141 at 16:53 UTC as b9850, registering t_layer_inp on qwen3next so drafters like DFlash v2 and EAGLE3 stop hitting a runtime assertion when they try to read target hidden states. qwen3next graphs had been failing to set res->t_layer_inp[il] since the model's initial contribution under time pressure from ModelBest during the qwen3 launch window. When Kashif Rasul landed DFlash v2 over the previous weekend, the drafter began requesting t_layer_inp on every layer of every model, exposing the qwen3next gap. jschmied's fix restores speculative decoding for qwen3next under both DFlash v2 and EAGLE3, including qwen-coder-next which shares the qwen3next backbone.

The rest of the queue: b9844 landed WebGPU NVFP4 support from yomaytk, b9848 fixed CUDA get_rows_back on tables with more than 65,535 rows, b9849 added IPv6-bracketed-authority parsing for llama-server URLs from ngxson, and Stanisław Szymczyk's b9851 fixed integer truncation in flash_attn_mask_to_KV_max on CUDA. Bolz and Szymczyk drove most of the Vulkan and CUDA work through the day.

Tech

Ollama 0.31.1 bumps MLX for small-batch matmul and improves Gemma 4 MTP

GitHub · 2026-06-30 · Source ↗ ·

Patrick Devine tagged 0.31.1 at 22:10 UTC. Jesse Gross tightened Gemma 4 MoE loading. Dan Hiltgen bumped MLX for Apple's new small-batch matmul kernel. Back-end llama.cpp lifted from b9822 to b9840, aligning with DFlash v2 and the Sunday CUDA work.

Patrick Devine cut Ollama v0.31.1 at 22:10 UTC on Tuesday. The release folds in three code changes and one dependency bump. The headline entry, credited to Jesse Gross, is a rewrite of the Gemma 4 mixture-of-experts loader that Ollama shipped in the 0.30 series. Gross tightened the code path so Gemma 4 MoE weights initialize without re-scanning the safetensors metadata on every fetch. Devine merged that as PR #16964. On the runtime side, Ollama bumps its embedded MLX to the current tip, which pulls in Apple's newly landed small-batch matmul kernel targeting the batch-size-1 to batch-size-8 regime where the previous kernel spent most of its time launching threads. Dan Hiltgen handled the MLX submodule update.

The llama.cpp back-end moves from b9822 to b9840, aligning Ollama's ggml layer with Kashif Rasul's DFlash-v2 sliding-window-per-layer-types work from the b9829 to b9835 window, ModelBest 张涛's MiniCPM5 chat parser, Adrien Gallouët's downloader --offline flag, and Sigbjørn Skjæret's --dump-prog Jinja debugger. Ollama's 0.31.1 changelog frames the llama.cpp update as a one-liner, but it inherits every ggml, CUDA and Vulkan improvement from the intervening 18 tag bumps, including Qualcomm's OpenCL flash-attention rewrite for Adreno and the cudaMemcpy2DAsync strided-copy fast path.

mlx: bump to latest version to include new small batch matmul kernel · improved gemma4 MTP performance.— Ollama 0.31.1 release notes

Downstream from the llama.cpp queue

Ollama's release notes describe Gemma 4 MTP performance as improved. That tracks the parallel work in llama.cpp b9847, where Johannes Gaessler fixed a CUDA regression that had disabled Gemma 4 E4B MTP entirely for weeks. The two changes together restore Gemma 4 speculative decoding for both Ollama's CUDA and MLX users on the same day. The nineteen release assets cover a standard matrix: Windows installer, macOS Darwin universal binary, Linux amd64 and arm64 with separate MLX-enabled and ROCm-enabled variants, and Jetson Jetpack builds.

tighten up gemma4 moe loading code.— Jesse Gross, PR #16964 title

The MLX-only builds now depend on the new small-batch matmul kernel to be present, so Devine bumped the minimum macOS build requirement in the release matrix. Devine, Hiltgen and Gross are the only three credited contributors. No new model families, no new API endpoints. The release is a clean maintenance tag that lands the June 22 through June 29 llama.cpp queue into the Ollama runtime and unblocks Gemma 4 MoE users who had reported slower cold-starts on the 0.30 series. Ollama's Web-UI and turbo-serving path are unaffected. 0.31.1 is scoped entirely to the local runtime.

Tech

LangGraph 1.2.7 patches DeltaChannel snapshot overwrite and langgraph-api UUIDs

GitHub · 2026-06-30 · Source ↗ ·

LangChain shipped 1.2.7 at 01:24 UTC with a DeltaChannel superstep snapshot-overwrite fix, valid UUIDs for exit-mode delta task IDs, LangSmith bumped 0.8.0 to 0.8.18, and a cluster of security dependency bumps.

The LangChain release bot tagged LangGraph 1.2.7 at 01:24 UTC on Tuesday, opening the June 30 news window. The commit bundles eleven changes across two bug fixes and nine dependency bumps. The first bug fix addresses a data-loss window in DeltaChannel where a snapshot taken mid-superstep could overwrite delta values that later steps depended on. LangGraph's DeltaChannel exists so that agents can commit incremental updates to shared state without full-graph rewrites, and a superstep pause was clobbering accumulated deltas whenever the checkpoint serializer intervened. The fix ensures Overwrite objects survive JSON round-trips, which had been silently downgrading to plain dicts and dropping their sentinel type on reload. LangChain pushed the fix through PR #8125 and PR #8127.

The second fix, PR #8165, changes how langgraph-api emits task IDs for exit-mode deltas. Previously the API surfaced integer counters or short strings that langgraph-api consumers had to disambiguate manually. The 1.2.7 fix emits proper UUIDs, aligning exit-mode task IDs with the format used everywhere else in the LangGraph state model. This matters for LangSmith trace collectors that key on task ID equality, and for third-party observability tools that had been miscounting exit-mode nodes as new tasks because their IDs looked truncated. LangSmith moves from 0.8.0 to 0.8.18 in the same release, so LangGraph and its trace back-end update together.

Fix snapshot DeltaChannel overwrite supersteps · Ensure Overwrite survives JSON roundtrips.— LangGraph 1.2.7 PR titles

Security-heavy dep bumps

The dependency bumps are heavy on security. Cryptography moves from 46.0.7 to 48.0.1, closing two CVEs in the underlying OpenSSL binding. PyJWT lands 2.13.0. Bleach, Jupyter Server, and Tornado all get security-only bumps as part of the Redis client update chain. The remaining PRs (#8176, #8164, #8107, #8106, #8134, #8108, #8092) all carry the fix(deps) prefix. Nothing in the release changes public LangGraph API surfaces. Agent authors upgrading from 1.2.6 should not need code changes.

Emit valid UUIDs for exit-mode delta task IDs for langgraph-api.— PR #8165 title

The LangGraph 1.x line has emphasized checkpoint durability across the 1.2 patch series. DeltaChannel is central to how LangGraph handles concurrent writes from parallel agents. The Overwrite JSON round-trip issue only surfaces when checkpoints move between Redis and a persistent SQL store, since the intermediate JSON encoding is where Overwrite's type metadata gets stripped. LangChain's release did not attribute the report, but the PR discussion references a langgraph-api production incident from the week prior in which delta values silently reverted to their pre-superstep state after a checkpoint restore. The 1.2.7 tag is the third patch on the 1.2 line in five days.

Tech

Pydantic AI 2.2.0 adds Sonnet 5, TwelveLabs Pegasus video, GoogleProvider retry_options

GitHub · 2026-07-01 · Source ↗ ·

Samuel Colvin's team cut 2.2.0 four minutes before midnight UTC with five features, sixteen bug fixes, and fifteen first-time contributors. Sonnet 5 lands as the first non-Anthropic agent framework identifier. TwelveLabs Pegasus becomes the first video-native tool example in the repo.

Pydantic AI v2.2.0 landed at 00:01 UTC on Wednesday, four minutes inside the June-30 news window and one hour after Anthropic's press embargo on Sonnet 5 lifted. The commit adds five new features, sixteen bug fixes, and introduces fifteen first-time contributors. The Claude Sonnet 5 entry, contributed by an Anthropic partner engineer, exposes the new claude-sonnet-5 model identifier through the Pydantic AI Anthropic provider and wires up adaptive-thinking so agent authors do not have to pass any special sampling arguments. Pydantic AI is the first agent framework outside of Claude Code itself with Sonnet 5 identifiers merged and tagged.

The TwelveLabs Pegasus example is the more novel addition. Mohit Chandel from TwelveLabs, one of the fifteen new contributors, added a video-understanding example that lets Pydantic AI agents ingest raw video content through Pegasus and reason over the model's outputs. Pegasus takes video and returns dense captions and temporal-grounded event tokens, and the example shows a Pydantic AI agent using those tokens as tool outputs. It is the first video-native tool example in the Pydantic AI repository. Colvin's team also added a factory-function argument to Dataset.evaluate, letting evaluation lifecycles instantiate providers per test case instead of once at the module level.

Claude Sonnet 5 (claude-sonnet-5) model support · TwelveLabs Pegasus video-understanding example integration.— Pydantic AI 2.2.0 release notes

Provider parity and sixteen bug fixes

OpenRouter now surfaces prompt and completions cost fields in usage metadata, closing a gap where OpenRouter customers could not audit per-request spend inside their evaluation pipeline. GoogleProvider gains a retry_options argument mirroring AnthropicProvider and OpenAIProvider, letting agents attach exponential-backoff retry policy directly to a provider instance rather than wrapping every call in an outer retry.

retry_options parameter added to GoogleProvider.— Pydantic AI 2.2.0 release notes

The sixteen bug fixes are unusually rich for a minor release. Image detail forwarding from vendor_metadata was silently dropping on GroqModel and MistralModel, so multimodal agents were passing lower-resolution images than expected. Sync callbacks were failing on Pyodide and emscripten because Pydantic AI was attempting to schedule them on a nonexistent event loop. output_audio_tokens was missing from usage tracking. Heroku model name routing was broken through family profiles. CodeExecutionTool was missing from GoogleModel's server-side invocation check. Groq reasoning tokens were not surfacing. There was a double-firing map join ID bug in workflow orchestration. A ClosedResourceError from graph run cancellation was surfacing to users instead of being suppressed. The fifteen first-time contributors include Mohit Chandel from TwelveLabs, immuhammadfurqan on OpenRouter cost fields, and tcconnally on the ClosedResourceError suppression. Colvin's release cadence has held at roughly one minor version every ten days across 2.0 and 2.1.

TechApplication

Willison ships shot-scraper 1.10 with a video subcommand for agent-recorded demos

Simon Willison · 2026-06-30 · Source ↗ ·

A new shot-scraper video command takes a storyboard.yml and drives Playwright 1.61 to record WebM demos of coding-agent work. Willison built it primarily with GPT-5.5 xhigh in Codex Desktop, and pins playwright-python 1.61.0 to unlock viewport-width customization.

Simon Willison published shot-scraper 1.10 on Tuesday afternoon, introducing a shot-scraper video subcommand that takes a storyboard.yml file and records a video of a Playwright-driven browser routine. The core idea, one Willison has argued for since early 2026, is that coding agents should hand back short video demos of their own work rather than screenshots or textual reproduction steps. The 1.10 release is his production-ready take on that pattern.

The command depends on Playwright's screencast mechanism, which Playwright shipped in the 1.59 line and stabilized in 1.61. Willison pins shot-scraper 1.10 to playwright-python 1.61.0 explicitly so that viewport-width customization works, since 1.60 had a regression in that path. The storyboard.yml format is validated by a Pydantic model and supports a list of actions per step: navigate, wait_for_selector, click, fill, screenshot_marker, sleep, and scroll. Each step can carry an optional annotation that appears as a caption overlay. Authentication uses a JSON cookie file, matching the format shot-scraper already accepts for its screenshot command. Output defaults to WebM. Willison added an --mp4 flag that transcodes through ffmpeg.

I've written before about the importance of having coding agents produce demos of their work; this is my latest attempt at enabling them to do that.— Simon Willison, on the shot-scraper video command

The agent-produced-demo argument

The blog post explicitly notes that GPT-5.5 xhigh in OpenAI's Codex Desktop wrote both the code and the documentation. Willison used the release as a small dogfooding exercise, treating the feature build itself as the kind of workflow that shot-scraper video is meant to record. He filed the original issue in February 2024 asking for a video-recording mode, but the tooling was not ready. Playwright's screencast in early 2024 could not customize viewport width, and Pydantic's model-driven YAML parsing was still awkward for Willison's target audience of coding agents. Both blockers cleared over the following two years.

I really like this pattern where the --help output for a command provides enough detail that a coding agent can use it.— Simon Willison

Willison explains his enthusiasm for the pattern by pointing at the ergonomics of the --help output. He argues that when an agent can read one command's --help output and understand exactly how to invoke it, the agent can chain that command into its own workflow without any human intervention. He wants shot-scraper video to be the canonical example of that pattern for agent-produced video demos. In his framing, a coding agent finishing a task should compose a small storyboard.yml, run shot-scraper video, and attach the WebM to its PR alongside the code changes. The release is small in code terms, but the pattern implications are larger: shot-scraper's install base already includes several coding-agent frameworks.