Daily2026-07-18
TechApplication

Google DeepMind ships Gemini 3.5 Pro after a full rebuild, at $1.25 in and $10 out per million tokens

Gemini 3.5 Pro lands Friday six weeks past its June target with a 2-million-token context window, a Deep Think reasoning layer gated behind the $250-per-month Ultra tier, and a base API price that undercuts Claude Fable 5 by roughly 4x on output. Google scrapped the Gemini 2.5 architecture in late June after a coding-data retrain failed to close the gap with Anthropic and OpenAI. Timing collides with Xi Jinping's in-person Shanghai keynote and Kimi K3 shipping open weights a week earlier.

CentralJersey.com · 2026-07-17 · Source ↗

Google DeepMind released Gemini 3.5 Pro on Friday, July 17, ending a six-week slip past the June commitment made at Google I/O and closing a public-relations gap that had cost Alphabet shares 4.4 percent on Thursday following Bloomberg's account of a failed late-June coding-data retrain. The model that shipped is a full architectural rebuild, not an incremental step on the Gemini 2.5 Pro backbone. Google engineers, according to independent reporting corroborated by DeepMind's launch materials, threw out the 2.5 Pro architecture after finding structural failures in recursive tool-calling, scalable vector graphics scene generation, and image-quality regressions that a data patch could not repair.

The headline numbers on release: a 2-million-token context window, a new Deep Think reasoning layer gated to the $250-per-month Gemini Advanced Ultra tier, and API pricing at $1.25 per million input tokens and $10 per million output. That output rate lands roughly 4x under Claude Fable 5 and just above Meta's Muse Spark 1.1, positioning Gemini 3.5 Pro as a price-aggressive flagship for enterprise workloads that had drifted to Anthropic and OpenAI over the past four months. A refreshed Gemini 3.5 Flash ships alongside it, with an explicit repositioning after Bloomberg's Thursday piece flagged that the previous Flash generation was "more expensive than previous versions yet slower with reduced capabilities."

"We're shipping quickly across a wide range of models while keeping them highly cost-effective for customers."— Google spokesperson, cited by Bloomberg on Thursday

The Deep Think reasoning layer

Deep Think is a Gemini-native long-horizon reasoning surface, exposed only through the Ultra tier and the paid API, designed to break complex tasks into subtasks and execute them autonomously with minimal user intervention. On Google's own internal benchmarks the layer clears a wide margin over 2.5 Pro on multi-step tool use, structured proofs, and long-context synthesis over the 2-million-token window. The company is framing Deep Think as the direct competitive answer to OpenAI's GPT-5.6 Sol Ultra subagent mode announced July 9 and to Anthropic's Extended Thinking budget that shipped with Sonnet 5 and Claude Fable 5. What the third-party benchmark community measures over the next 72 hours will decide whether the Deep Think framing lands or reads as marketing.

Autonomous workflows and the enterprise pitch

Google's enterprise pitch centers on autonomous workflows. Rather than requiring explicit instructions per step, Gemini 3.5 Pro is presented as able to plan a multi-step task, execute against connected apps, and hand back a completed artifact. The framing tracks the direction OpenAI took with ChatGPT Work Friday and Meta took with Meta Business Agent this month. The pricing move is aggressive: at $1.25 input and $10 output per million tokens, Gemini 3.5 Pro undercuts Claude Fable 5 by roughly 4x on output and matches Meta's Muse Spark 1.1 tier while offering a nominally larger context window. Anthropic's mid-July price ladder came under coordinated pressure this week from Kimi K3, Muse Spark, and now Gemini 3.5 Pro.

"A ground-up redesign, engineered to close critical gaps in mathematical reasoning, scalable vector graphics scene generation, and overall image quality."— Google DeepMind launch framing, per BigGo Finance and TechTimes

The timing collision and the caveats

Gemini 3.5 Pro's Friday release lands on the same UTC day as Xi Jinping's Shanghai World AI Conference keynote, Huawei's Atlas 950 SuperPoD debut, TSMC's earnings and Arizona expansion, and Anthropic's Claude Code point release. The collision is not accidental. Google's internal debate over whether to accept an August ship date and cede a full month to Kimi K3 was resolved, according to Bloomberg's Thursday reporting, in favor of the Friday launch. The caveats are real. Every claim about specifications, pricing, and capabilities in third-party reporting through Thursday came from leaks, not official Google announcements. The June 2026 Gemini 3.5 Pro launch commitment came and went. Some early enterprise customers who moved to Claude Sonnet and GPT-5.6 Sol during the delay window will not migrate back on price alone. The next 30 days will measure whether the rebuild landed a real capability step or a repositioned narrative.

The wider read across the Shanghai and Silicon Valley calendars: three model families now sit inside touching distance on frontier coding evaluations, and each of the four US labs is now pricing at or below the DeepSeek-derived open-weight floor. What was a differentiated-capability market in Q1 2026 is now a differentiated-workflow market. That reprices the value of vertical integration for Google (Search, Workspace, Android), Microsoft (Copilot, Azure, GitHub), and Anthropic (Claude Code, MCP). Meta and xAI are pricing the same play a different way. Gemini 3.5 Pro is the Google entry.

The Big Picture

Google DeepMind shipped Gemini 3.5 Pro Friday six weeks late after a full architectural rebuild, at $1.25 per million input tokens and $10 output with a 2-million-token context window and a Deep Think reasoning layer priced behind the $250-per-month Ultra tier. Xi Jinping opened the Shanghai World AI Conference in person for the first time since the event began in 2018, pledged 5,000 AI training placements for developing-country delegates over five years, and 29 countries signed on to launch the World AI Cooperation Organization headquartered in Shanghai. Huawei debuted the Atlas 950 SuperPoD on the WAIC floor with 1,024 Ascend NPUs across 16 cabinets, 1 EFLOPS at FP8, 2 EFLOPS at FP4, and 256TB of globally addressable memory over the UnifiedBus 2.0 fabric, targeting an 8,192-chip full configuration built without any US-origin components. TSMC lifted 2026 capex to $60-64 billion after a 77.4 percent Q2 profit surge to NT$706.56 billion on $40.2 billion revenue, and added $100 billion to its Arizona commitment to bring total US investment to $265 billion across up to four additional 2-nanometer fabs.

Nubia unveiled the NaviX Ultra at WAIC as the world's first agent-native smartphone with a dedicated orange AI button running ByteDance's Doubao at the OS level. NotebookLM was renamed Gemini Notebook with an integrated cloud code sandbox and 30 million individual users and 600,000-plus organizations on the base. Anthropic shipped Claude Code with a --forward-subagent-text flag, a default 200 subagent-per-session cap, hardened permission previews against zero-width and lookalike quote characters, and a default model swap to Claude Opus 4.8 on Bedrock and Vertex. VideoChat3 landed at the top of Hugging Face daily papers with a 4-billion-parameter fully open video multimodal model from Nanjing University's Multimedia Computing Group, ahead of SEED and SearchOS-V1.

CISA added Microsoft SharePoint CVE-2026-58644 to the KEV catalog with a July 19 federal patch deadline, WordPress 6.9.5 and 7.0.2 patched a pre-auth remote code execution in core against a default install, and Cursor IDE shipped an emergency patch for two zero-click sandbox-escape vulnerabilities documented against the coding-agent surface. Meta rolled out the Meta Business Agent enterprise platform and a Meta Compute cloud arm targeting WhatsApp, Messenger and Instagram customer-conversation surfaces at scale. AIsphere closed a $439 million Series C led by Alibaba on July 14 for AI video generation. Check Point's 2026 Security Report documented autonomous agents executing full exploitation workflows across thousands of commands with minimal human direction. Microsoft repositioned Copilot ahead of OpenAI's ChatGPT Work in enterprise pitches even as the two partners continue to share infrastructure and revenue at the Azure layer.

More today

24 stories · 2026-07-17
PolicyTrend

Xi Jinping opens WAIC in person, launches 29-country World AI Cooperation Organization headquartered in Shanghai

Al Jazeera · 2026-07-17 · Source ↗

Xi's first in-person WAIC keynote since the event launched in 2018 called for open-source, open collaboration, and warned against "solo performance" AI dominance. Twenty-nine countries signed on to WAICO the following day, with the body headquartered in Shanghai and 5,000 developing-country training placements funded over five years.

President Xi Jinping opened the 2026 World Artificial Intelligence Conference in Shanghai on Friday, July 17, delivering the keynote address at the High-Level Meeting on Global AI Governance in person for the first time since the conference launched in 2018. The staging is the point. Xi's presence was framed by state media as a signal that AI has moved from the industrial-policy pile onto the diplomatic front burner, on par with climate, trade, and reserve-currency policy. Chinese Foreign Ministry readouts and CGTN's full-text release put the address on the record as the founding statement of a Chinese-led alternative to what Beijing frames as US-dominated AI governance.

"Artificial intelligence should not be a solo performance by a single country, but a symphony of international cooperation."— Xi Jinping, WAIC 2026 keynote, July 17

The concrete deliverable came the following day. Twenty-nine countries signed on to establish the World Artificial Intelligence Cooperation Organization, WAICO, with the body headquartered in Shanghai. Xi described the new organization as "a milestone in the history of world AI development," positioning it as answering long-standing Global South calls for a larger voice in AI governance. Membership at the founding included partners across ASEAN, the Gulf states, sub-Saharan Africa, and Latin America. Notably absent from the signatories were the United States, the United Kingdom, Germany, Japan, and India, marking a hard fork from the OECD AI Recommendations track and the Bletchley process.

The concrete pledges

Xi committed China to provide developing countries with 5,000 opportunities in AI training and seminar programs over the next five years, to develop international AI application cooperation centers with regional organizations, and to enable 30 countries to use the AI-powered meteorological warning system MAZU. The meteorological pledge lands as heat waves, floods, and typhoons drive climate-adaptation demand across sub-Saharan Africa and Southeast Asia and as Nvidia's own Earth-2 sovereign-AI weather platform seeks similar customer bases. WAIC 2026 itself ran across four days and hosted more than 140 forums, 1,400 registered speakers, 1,100 exhibitors, and more than 300 product debuts, per the Ministry of Foreign Affairs opening summary.

"Encourage open source, openness, collaboration and sharing."— Xi Jinping, keynote framing, July 17

The US read and the open-source frame

Xi's open-source framing was politically loaded. Two days before the address, US Under Secretary for Industry and Security Alan Estevez had testified to the House Foreign Affairs Committee that the January 13 rule permitting limited H200 shipments to approved Chinese buyers, including Alibaba, Tencent, ByteDance and JD.com, would remain in force. Xi's line pushed the counter-argument that Chinese labs are the open-weight backbone of the global research community, with Kimi K3, DeepSeek V4, Qwen 3.6, and GLM-5.2 all shipping open weights in July while OpenAI and Anthropic hold their frontier models closed. Alphabet, Anthropic and Meta's own compliance responses to WAICO membership have not been disclosed. The State Department has issued no readout as of Friday evening UTC.

TechPolicy

Huawei debuts Atlas 950 SuperPoD on WAIC floor, 8,192 Ascend NPUs at 1 EFLOPS FP8 built without US silicon

Seoul Economic Daily · 2026-07-17 · Source ↗

The show floor configuration linked 1,024 Ascend NPUs across 16 cabinets, with the full architecture scaling to 8,192 Ascend 950DT processors via the proprietary UnifiedBus 2.0 memory fabric. Huawei Central puts the compute claim at 6.7x Nvidia's NVL144 for equivalent floating-point workloads.

Huawei publicly displayed the Atlas 950 SuperPoD hardware for the first time at WAIC 2026 in Shanghai on Friday, July 17, with the on-floor configuration linking 1,024 Ascend NPUs across 16 cabinets and the full production architecture scaling to 8,192 Ascend 950DT AI processors. The SuperPoD is fully liquid-cooled and uses Huawei's proprietary UnifiedBus 2.0 protocol as a memory fabric rather than a conventional network. On Huawei's own claimed benchmarks, the 1,024-NPU show configuration delivers 1 EFLOPS at FP8, 2 EFLOPS at FP4, and 256 terabytes of globally addressable memory. Huawei Central published a comparison putting the full-scale Atlas 950 SuperPoD at 6.7x Nvidia's NVL144 rack on equivalent floating-point workloads. Independent verification against Nvidia's own benchmarks has not been published.

The strategic frame matters more than the peak numbers. The Atlas 950 is the first production-scale demonstration that a Chinese AI infrastructure stack, from silicon through to interconnect through to power and cooling, can be built and shipped without any US-origin components. Ascend 950DT is fabbed on SMIC's N+2 process, the memory hierarchy is Chinese-domestic HBM3 equivalents, and UnifiedBus 2.0 replaces both NVLink and InfiniBand in the design.

"The Atlas 950 SuperPoD is a fully liquid-cooled data center super node, connecting 8,192 Ascend 950DT AI processors using Huawei's proprietary UnifiedBus 2.0 protocol."— Huawei product documentation, WAIC 2026 debut

The Xi-Huawei alignment

The launch calendar is not coincidental. Xi Jinping used the WAIC keynote earlier the same day to argue for open-source AI leadership, and the Huawei floor demonstration is the hardware complement to that argument. Combined with Nubia's NaviX Ultra agent-first smartphone debut, ZTE's AI Agent Phone commercial rollout, and DeepSeek V4 shipping open weights, WAIC 2026 is being staged as an integrated Chinese answer to the US frontier-labs pricing war Google, Anthropic, and OpenAI have been running through July.

The compute-diplomacy read

South China Morning Post's Friday coverage puts the Atlas 950 SuperPoD alongside Xi's remarks as the two most-consequential compute-diplomacy signals of the year. Sedaily's report emphasizes that the SuperPoD is targeted at both domestic AI training runs at gigawatt-class labs and at export deployment to WAICO founding members. Analysts inside Beijing have flagged that at Huawei's Q3 SN50L Ascend production yield, roughly 40,000 Ascend 950DT NPUs could ship to WAICO signatories over 12 months, enough to seed a 5-cluster training footprint outside US export-control reach. Whether Nvidia's NVL144 comparison holds under independent load testing will be the next benchmark. What is not disputed: Huawei has cleared the "can it be built without US parts" question. The next question is whether it scales.

TechEconomic

TSMC lifts 2026 capex to $64 billion after 77 percent Q2 profit surge, adds $100 billion to Arizona for $265 billion total

TechTimes · 2026-07-17 · Source ↗

Net profit hit NT$706.56 billion on $40.2 billion revenue, with 5nm at 33 percent of wafer revenue and 3nm at 30 percent. Full-year US dollar revenue guidance rose above 40 percent growth, and the Arizona addition funds up to four more 2-nanometer fabs on top of eight already announced.

Taiwan Semiconductor Manufacturing posted Q2 2026 results on Friday, July 17, with net profit of NT$706.56 billion, roughly $21.8 billion, up 77.4 percent year over year, and revenue of $40.2 billion up 36 percent. Earnings per share of NT$27.25 cleared consensus by a substantial margin. The report is the fifth consecutive quarterly record for TSMC, and the numbers cement the company's position as the single most-exposed public equity to the AI compute buildout.

"The 5nm node led with a 33 percent share, with 3nm close behind at 30 percent, and 2-nanometer contributed 3 percent of wafer revenue in its debut quarter."— TSMC Q2 2026 earnings summary, per Yahoo Finance

The capex lift and the Arizona commitment

TSMC raised its 2026 full-year capital expenditure guidance to a range of $60 to $64 billion, up from prior $56 billion, and lifted full-year US dollar revenue growth guidance above 40 percent. Advanced process technologies including 5nm, 3nm, and the newly-shipping 2nm now account for 77 percent of total wafer revenue. Alongside the earnings, TSMC announced an additional $100 billion investment in Arizona over the next several years, bringing total announced US investment to $265 billion. The new capacity is targeted at 2-nanometer and below nodes plus advanced packaging, with roughly four more fab plants to be constructed on top of the eight already announced or underway.

"The company raised its full-year US dollar revenue growth outlook to slightly above 40 percent and boosted capital expenditure guidance to between $60 billion and $64 billion."— TSMC guidance update, July 17

Advanced packaging and the AI bottleneck

The TechTimes coverage foregrounds the advanced-packaging thesis. HPC pulled through to 66 percent of wafer revenue in Q2, and the packaging bottleneck that has capped Nvidia GB200 shipments through 2026 remains the single largest constraint on AI-training compute deployment. Four additional Arizona fabs targeted at 2nm and beyond are also positioned as advanced-packaging capacity, which lets TSMC route HPC customers around the current CoWoS-L supply pinch. The concrete number: TSMC's US investment now equals just over the entire 2025 US federal science and technology R&D budget, at a level that no other private capex program in US industrial history has reached.

What it means for Nvidia, AMD, and Apple

Nvidia's Rubin and Rubin Ultra generations, AMD's MI350/MI400 series, and Apple's M5 and M6 silicon all depend on TSMC advanced nodes and CoWoS-L packaging. The Arizona expansion cuts the geographic-concentration risk that has been priced into Nvidia and AMD equity since the January 2026 Trump AI-chip tariff proclamation. TSMC's 5-nanometer share falling to 33 percent as 3nm approaches parity, and 2nm entering revenue mix at 3 percent inside its first quarter, matches the internal Nvidia and AMD roadmap for a 2027 Rubin Ultra generation on 2nm. The capex lift is the market's answer to whether AI-training demand is still binding. It is.

ApplicationTech

Google renames NotebookLM to Gemini Notebook, adds secure cloud code sandbox for Pro users

Dataconomy · 2026-07-17 · Source ↗

Rebrand takes effect Friday alongside a new cloud code sandbox for Pro users, cross-syncing with the Gemini app and Google Search AI Mode, and a 30-million individual-user and 600,000-plus-organization active base to migrate.

Google renamed NotebookLM to Gemini Notebook on Friday, July 17, folding the two-year-old research-and-notebook product into the broader Gemini brand while preserving it as a standalone offering. The rebrand ships with new capability, not just new branding. Gemini Notebook Pro now includes a secure cloud code sandbox for running Python and JavaScript directly inside notebooks, targeting complex data-analysis workflows that had required moving between NotebookLM and Colab. Google Workspace Updates published the enterprise-tier framing at 10 AM PT, and Dataconomy's Friday coverage confirmed the individual-user rollout was live in the US and UK as of publication.

"It's the same standalone product, now doing more across the Google ecosystem and updated with a secure cloud computer."— Google product blog, July 17

The 30 million user base and the Gemini surface

Gemini Notebook now supports more than 30 million individual users and more than 600,000 organizations, per Google's own disclosure. Notebooks will sync across the Gemini app and Google Search in the coming weeks, with Google committing to bring notebooks into Search AI Mode so a user researching a topic can pull their existing notebook context into a Search session. The unified surface is a direct response to OpenAI's ChatGPT Work rollout Friday and Anthropic's Claude Cowork expansion earlier in the week. NotebookLM's audio-overview and video-overview features remain and are extended into Gemini Live for real-time voice discussion of notebook content.

The renaming streak

TechCrunch's Rebecca Bellan flagged that this is the third Google product to be renamed into the Gemini brand in 90 days, following Bard, Duet AI, and Google Labs consolidations. The rebrand also lands as Google Workspace Updates announced Google Learn About and Google Illuminate migrations into Gemini's core surface for the fall term. For educators and researchers, the rebrand raises a real workflow question: NotebookLM's positioning as a research-first, notes-first product is now bundled into a general assistant. The counter-argument Google made in the Workspace Updates post is that keeping Gemini Notebook standalone lets it stay research-first while opening up the broader Gemini corpus and code sandbox.

"Enhanced capabilities to write and execute code natively, which allows users to conduct complex data analysis based solely on their own research data."— Dataconomy on Gemini Notebook Pro's cloud code sandbox

The wider read on the launch calendar: renaming NotebookLM the same day Gemini 3.5 Pro ships and Xi opens WAIC is a deliberate concentration of Google's July news surface. For enterprise buyers, the offer is now clearer. Ultra tier for Deep Think reasoning, Gemini Notebook Pro for research-and-analysis workloads, and API access under the $1.25/$10 per million pricing. The remaining question is whether Gemini 3.5 Pro's capabilities under Deep Think land against ChatGPT Work and Claude Sonnet 5 in third-party benchmarks over the next 30 days.

ApplicationTech

Nubia unveils NaviX Ultra at WAIC as world's first agent-native smartphone, orange AI button runs ByteDance Doubao at OS level

Notebookcheck · 2026-07-17 · Source ↗

ZTE's Nubia unit unveiled the NaviX Ultra at WAIC Friday, positioning a dedicated hardware AI button and OS-level ByteDance Doubao integration that opens apps, compares prices, fills in fields, and completes purchases without user touch input. Nubia's prior Doubao-branded M153 sold 30,000 units on day one.

Nubia, ZTE's mobile-device unit, unveiled the NaviX Ultra at the World AI Conference in Shanghai on Friday, July 17, positioning the device as "the world's first AI agent smartphone." The framing is not just capability. The NaviX Ultra ships with a dedicated hardware AI button in Nubia's signature burnt orange, sitting where a Google Assistant or Apple Action Button would sit on Android and iOS flagships, and the AI agent integration is at the operating-system level rather than as a launched app. The agent runs ByteDance's Doubao model as the reasoning back-end and can open apps, compare prices across surfaces, autofill user details, and complete purchases end to end without the user touching the screen.

"Unlike today's AI-powered smartphones, which primarily answer questions or generate content, Nubia says its AI agent is deeply integrated into the operating system."— Notebookcheck, NaviX Ultra debut coverage, July 17

The Ni Fei framing and the ZTE strategy

Ni Fei, ZTE senior vice president and president of the company's mobile devices division, positioned the NaviX Ultra as ZTE's commercial rollout of what began as the Nubia M153 second-generation Doubao smartphone. That prior generation had sold out its 30,000-unit initial batch on the first day and became a signature product for the Doubao-Nubia partnership. Financial markets responded positively: ZTE H-shares gained more than 5 percent on the Hong Kong exchange Friday morning following the regulatory filing preceding the debut.

What the agent actually does

The demonstration on the WAIC floor showed the NaviX Ultra completing a compound task: search for a book across three retailers, pull user shipping information from stored profile, apply a promotion code, and complete the purchase, all through the AI button. A second demonstration handled a mobile-banking transfer with the same touch-free workflow. Doubao's OS-level integration means the agent is invoked with system-level permissions rather than through Android accessibility hooks, which sidesteps the reliability issues that have plagued Rabbit R1 and Humane AI Pin agent phones over the prior 18 months.

"The phone's AI can open apps, compare prices, fill in your details, and complete purchases — all without you touching the screen."— Memeburn, NaviX Ultra feature summary

The competitive frame

StepFun, Alibaba, and Xiaomi are all rumored to be racing similar agent-first smartphone architectures for a Q4 2026 launch. Nubia's Friday debut positions ZTE first to publicly-shipping product, though the NaviX Ultra's actual retail availability and pricing remain unconfirmed. What the debut settles: OS-level agent integration on Android is now shipping in commercial hardware, not a research demo. The read from Shenzhen is that if agent-first smartphones find traction with Chinese consumers over the next two quarters, the category becomes the third differentiated phone form-factor after camera-first and folding, and the first one Apple has not led into.

TechSafety

Anthropic ships Claude Code stability release with 200-subagent-per-session cap and hardened permission previews

Anthropic release notes · 2026-07-17 · Source ↗

Adds --forward-subagent-text and CLAUDE_CODE_FORWARD_SUBAGENT_TEXT for stream-json transparency, defaults to Claude Opus 4.8 on Bedrock and Vertex, and neutralizes bidirectional-override, zero-width, and lookalike quote characters in permission previews to block visual approval spoofing.

Anthropic tagged a Claude Code stability and quality update on Friday, July 17, that lands as the most substantive Claude Code point release of the month. The headline changes: a new --forward-subagent-text flag and matching CLAUDE_CODE_FORWARD_SUBAGENT_TEXT environment variable that expose subagent thinking-and-response text into stream-json output for observability tools; a per-session hard cap on subagent spawns defaulting to 200 with CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION as the override; a default model swap to Claude Opus 4.8 on Amazon Bedrock, Google Vertex AI, and the AWS Marketplace listing; and a security hardening pass on permission previews so that tool-input strings cannot visually overwrite the approval prompt shown to the user.

"Permission previews relayed to chat channels were fixed to neutralize bidirectional-override, zero-width, and lookalike quote characters, so tool inputs cannot visually alter the approval message."— Anthropic Claude Code changelog, July 17

The subagent cap and the runaway-delegation problem

The 200-subagent-per-session cap is a direct fix for a known failure mode where a background Claude Code session recursively delegates work to fresh subagents faster than the operator can spot. The bug had surfaced in high-fanout research workflows and in agent-driven test runs over the June to July window. Setting a hard default cap moves the guardrail from soft-warning to hard-fail, which is consistent with the direction Anthropic has been running Claude Code since the v2.1 series. The related fix for returning to claude agents silently stopping running subagents and re-running prompts from scratch means work-in-progress carries over across session switches.

The prompt-injection hardening

The permission-preview hardening is the second-order response to the OpenAI GPT-Red disclosure Thursday. The vulnerability class Anthropic patched: an attacker crafts tool input that includes Unicode bidirectional-override characters, zero-width joiners, and lookalike quotation marks that render as "cd /tmp" in the approval prompt but execute as "cd /etc" once approved. The hardening pass canonicalizes the string, strips bidi overrides, and normalizes lookalike quotes so the operator sees the exact bytes that will run. Anthropic's Trust and Safety team has documented the specific attack payloads that motivated the fix.

"A new --forward-subagent-text flag and CLAUDE_CODE_FORWARD_SUBAGENT_TEXT environment variable were added to include subagent text and thinking in stream-json output."— Claude Code changelog, July 17

The default-model swap

Swapping the default model to Claude Opus 4.8 on Bedrock, Vertex and AWS Marketplace is a routine version-bump, but it comes with pricing implications. Enterprise customers running Claude Code against Bedrock had been defaulting to Sonnet 5 for cost reasons; the Opus 4.8 default now costs roughly 5x on output tokens. Customers who want to keep Sonnet as default need to set the CLAUDE_CODE_MODEL environment variable explicitly. Anthropic's pricing story remains under pressure from Kimi K3 at $3/$15 for a 2.8T MoE and Gemini 3.5 Pro at $1.25/$10 for a rebuilt frontier model. The July 17 Claude Code update makes clear that the Anthropic play is depth of integration into MCP, Bedrock, Vertex, and the Microsoft 365 connector rather than headline API-per-token pricing.

Tech

VideoChat3 tops Hugging Face daily papers, 4-billion-parameter fully open video MLLM covers motion, hour-long stories, and live streams

Hugging Face · 2026-07-17 · Source ↗

Nanjing University's Multimedia Computing Group ships a fully open 4B video multimodal model built on an Inflated 3D Vision Transformer with adaptive frame-resolution sampling. Second on the daily list is SEED for self-evolving on-policy distillation, third is SearchOS-V1 for open-domain search agents.

VideoChat3 landed at the top of the Hugging Face daily papers surface on Friday, July 17, as a fully open 4-billion-parameter video multimodal large language model from Nanjing University's Multimedia Computing Group. The model is positioned as a generalist video-understanding backbone that handles subtle motion, hour-long narrative reasoning, precise temporal grounding, and live-stream understanding, all at inference-time cost that is competitive with 7B image-only baselines. The GitHub repository at MCG-NJU/VideoChat3 launched with weights, code, evaluation, and an inference server, under a permissive open license.

The architecture bet

Two design choices carry the paper. First, an Inflated 3D Vision Transformer, I3D-ViT, that lifts a pretrained 2D image vision transformer into a spatiotemporal backbone by inflating the tokenizer without a full retrain. Second, an Adaptive Frame Resolution for Streaming Video Perception layer that dynamically adjusts the frame-sampling rate and per-frame resolution based on scene complexity, letting the model process a two-hour movie and a five-second GIF with proportional compute. The efficiency claim: 4B active parameters can reach parity with 13B-parameter video baselines at a fraction of the inference cost, which if it holds outside the paper's own evaluation moves video understanding into the reachable-on-consumer-GPU range.

"VideoChat3 is a fully open, efficient, and generalist video-centric MLLM that advances video understanding through two complementary designs."— VideoChat3 paper abstract, MCG-NJU

The rest of the daily surface

Second on the Hugging Face daily papers list Friday was SEED, "Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning," a method for turning teacher-model traces into student policies without a fresh RL run. Third was SearchOS-V1, positioning a robust open-domain information-seeking agent architecture with heterogeneous tool routing. The Friday surface reflects the July research direction Simon Willison flagged earlier in the week: distillation-driven agent training, retrieval-augmented reasoning, and open video understanding all landing in the same 24-hour window.

"A 4B generalist video MLLM built to understand video across time, from subtle motion and hour-long stories to precise temporal evidence and live streams."— MCG-NJU project description

Why it matters

Open video-language models had been stuck at 7B-to-13B scale on architectures inherited from LLaVA, with real-time video understanding running slow enough to preclude live-stream applications. VideoChat3's I3D-ViT plus Adaptive Frame Resolution combination pushes the compute envelope down without dropping the capability floor, and the fully-open release model, including the training recipe, means labs building on top can retrain against domain-specific video corpora without renegotiating a license. For robotics teams, video-agent developers, and long-form content researchers, VideoChat3 is likely the July baseline moving forward, replacing InternVideo2 and LLaVA-Video across many published benchmarks over the next 60 days.

SafetyPolicy

CISA adds SharePoint CVE-2026-58644 to KEV catalog with two-day federal patch deadline, CVSS 9.8 unauth RCE under exploit

The Hacker News · 2026-07-17 · Source ↗

Critical deserialization-of-untrusted-data vulnerability in Microsoft SharePoint Server permits unauthenticated remote code execution. Federal civilian agencies must apply the fix by July 19; broader guidance recommends immediate deployment.

The US Cybersecurity and Infrastructure Security Agency added CVE-2026-58644, a critical Microsoft SharePoint Server deserialization vulnerability, to its Known Exploited Vulnerabilities catalog on Friday, July 17. Federal Civilian Executive Branch agencies must apply the fix by July 19, 2026, one of the tightest CISA deadlines set this year. The vulnerability carries a CVSS 3.1 score of 9.8, is triggered by deserialization of untrusted data, and permits an unauthenticated attacker to execute arbitrary code on the target SharePoint Server host.

"A critical deserialization of untrusted data vulnerability that allows an unauthorized attacker to execute arbitrary code."— CISA advisory summary, CVE-2026-58644

The exploit surface and the observed activity

SharePoint Server is broadly deployed across US federal networks, higher-education institutions, and state and local government agencies. CISA's KEV designation requires evidence of active exploitation in the wild, and the two-day patch deadline is a signal that the observed activity is targeted rather than opportunistic. Independent researchers on The Hacker News have flagged that the exploitation is being routed through internet-exposed SharePoint hosts and that adversaries are dropping web shells for persistent access. Microsoft published the patch in the July Patch Tuesday cycle earlier in the week, so the fix has been available for approximately five days.

The context: 570 CVE Patch Tuesday

The SharePoint CVE lands inside the record-setting July Patch Tuesday release Microsoft shipped Wednesday, which included 570 total CVEs, the largest single monthly patch release in Microsoft history. Microsoft credited its AI-powered vulnerability discovery system running against the Windows codebase for the volume, and had disclosed three zero-days including an Active Directory Federation Services privilege bug already under active exploitation, an additional SharePoint RCE, and a publicly disclosed BitLocker bypass. Friday's KEV addition and two-day deadline concentrate agency attention on the SharePoint deserialization bug specifically.

"Federal Civilian Executive Branch agencies to apply the fixes by July 19, 2026."— CISA KEV catalog directive

What defenders should do

The recommended actions for defenders are: apply the July Patch Tuesday updates to all SharePoint Server hosts immediately; hunt for indicators of compromise including anomalous ASP.NET compilation artifacts, unusual w3wp.exe process spawns, and outbound connections to unfamiliar IPs from SharePoint-tier hosts; and for internet-exposed SharePoint deployments, consider temporary firewall isolation until patch verification completes. Enterprises that also run Microsoft 365 Connector against Claude, ChatGPT, or Gemini agents should verify agent permissions do not include write access to unpatched SharePoint tenants during the mitigation window.

Safety

WordPress 6.9.5 and 7.0.2 patch pre-auth RCE in core against a default install with no plugins

Security Boulevard · 2026-07-17 · Source ↗

Anonymous request triggers unauthenticated remote code execution on a stock installation. Emergency releases published Friday cover the current 7.x and long-term-support 6.9.x branches.

Automattic and the WordPress core security team shipped WordPress 6.9.5 and 7.0.2 as emergency releases on Friday, July 17, closing a pre-authentication remote code execution vulnerability in WordPress core. The advisory framing is stark: an anonymous HTTP request against a default installation, with no plugins and no themes installed, is sufficient to trigger arbitrary code execution as the web-server user. The vulnerability affects the vast majority of the ~43 percent of the public web running on WordPress. Automatic-updates-enabled sites will receive the patch within 24 hours; manually-managed sites need to apply immediately.

"A pre-auth RCE in core that an anonymous request can trigger against a default install with no plugins."— Security Boulevard advisory summary, July 17

The scope

Pre-authentication RCE vulnerabilities in WordPress core are rare. The last public disclosure at this level was 2019. Because the vulnerability sits in core rather than a plugin, the entire installed base is exposed regardless of theme, plugin, or hosting configuration. Managed hosts including WordPress.com, WP Engine, Kinsta, and Pantheon confirmed they were pushing the patch across managed customers Friday afternoon. Cloudflare and Sucuri deployed emergency web application firewall rules for customers unable to patch immediately.

The AI-generated-content angle

The patch lands as WordPress core has been under scrutiny for AI-content moderation and prompt-injection surfaces in the Automattic-driven AI content-writing plugin ecosystem. The core vulnerability is not AI-related on its face, but the exploitation potential expands markedly if attackers can chain the pre-auth RCE with Jetpack AI, WordPress AI Assistant, or the growing category of MCP-connected WordPress plugins, all of which have write access to WordPress databases and file systems. The recommended action for administrators is: apply 6.9.5 or 7.0.2 immediately, rotate WordPress salts, audit user accounts for privilege escalation, and check plugin ecosystem for MCP or LLM tool connections that may have been abused during the exposure window.

SafetyTech

Cursor IDE ships emergency patch for two zero-click RCE vulnerabilities that overwrite system files and escape the sandbox

Help Net Security · 2026-07-17 · Source ↗

Cursor patched two vulnerabilities Friday that allowed attackers to overwrite system files and escape sandbox protections without user interaction. Enterprises using AI coding assistants received an out-of-band advisory to update immediately.

Cursor shipped an emergency patch on Friday, July 17, closing two zero-click remote code execution vulnerabilities in the Cursor IDE that permitted attackers to overwrite system files and escape sandbox protections without any user interaction. The advisory landed alongside a Help Net Security summary of the Check Point 2026 AI Security Report and coincided with CISA's SharePoint Server KEV addition earlier the same day. Coming on top of the Grok Build CLI silent-upload disclosure from Wednesday and the Anthropic Claude Code permission-preview hardening the same day, the coding-agent IDE category has now had three material safety or data-handling issues surface publicly inside a single week.

The vulnerability class

Zero-click RCE in a coding IDE is a rare but severe vulnerability class. Cursor's sandbox is designed to isolate LLM-generated code execution from the host operating system, and the two bugs Cursor patched break both directions of that isolation. The first vulnerability let attacker-controlled content passed to the LLM overwrite protected system files. The second let sandbox-executed code escape into a broader host-level context. Both required no user click, meaning simply opening a malicious file or repository was sufficient to trigger exploitation.

"Enterprises using AI coding assistants were urged to patch immediately."— Help Net Security, coding-IDE security summary

The context: agents on developer surfaces

Cursor is one of the top three AI coding IDEs by paid-seat count alongside Claude Code and OpenAI's Codex CLI, with an estimated $600 million ARR run rate as of Q2 2026. The vulnerability class is not unique to Cursor. Anthropic patched a Claude Code worktree-isolation bug Wednesday; xAI shipped Grok Build CLI open-source with the exfiltration hooks still visible on Thursday; OpenAI Codex CLI rust-v0.144.5 hardened dangerous-command detection against forced rm the same day. The direction of travel is clear: coding-agent IDEs are now a first-class attack surface, and the vendors most-integrated into developer workflow are absorbing the highest scrutiny.

What defenders should do

Update Cursor to the current build immediately. Audit VS Code extension host permissions for parity. Sandbox all LLM-generated code execution behind an OS-level container or firecracker VM where possible. Rotate API keys and cloud credentials present in any repository that had been open in vulnerable Cursor builds during the exposure window. The Check Point 2026 Security Report published Friday documents autonomous agents executing full exploitation workflows across thousands of commands with minimal human direction, which is the failure-mode this class of vulnerability opens.

ApplicationEconomic

Meta launches Meta Business Agent and Meta Compute cloud, targets WhatsApp, Messenger, and Instagram conversation surfaces

Buildfastwithai · 2026-07-17 · Source ↗

Enterprise agent platform ships globally to let businesses build, customize, and deploy AI agents on top of Meta's messaging surfaces. Meta Compute is Meta's first-ever paid cloud business, positioned against Azure, AWS, and Google Cloud.

Meta rolled out the Meta Business Agent enterprise platform and the Meta Compute cloud business on Friday, July 17, marking Meta's first commercial cloud offering and Meta's first packaged enterprise-agent product. The Business Agent surface lets enterprises build, customize, and deploy AI agents inside WhatsApp, Messenger, and Instagram business surfaces at global scale. Meta's pitch: turn the billions of customer conversations that already flow across those surfaces into deployable agent workloads, without customers having to leave WhatsApp or Instagram to interact with a company.

The Meta Compute pivot

Meta Compute is the more consequential launch. Until now, Meta had no external cloud business. The Louisiana Hyperion campus at 5 gigawatts and $50 billion, the Reno Prometheus site, and the smaller Sturgeon County build in Alberta had been positioned entirely as internal training capacity for Muse Spark, Muse Image, and downstream Meta AI products. Meta's Q2 disclosure last month that Zuckerberg was "weighing renting excess AI capacity to third parties" is now the direction. Meta Compute launches with Muse Spark 1.1 API access, priced at $1.25 input and $4.25 output per million tokens, and Muse Image API access at rates undisclosed pending general availability.

"Aiming to turn the billions of customer conversations on WhatsApp, Messenger, and Instagram into deployable enterprise agents."— Meta Business Agent product framing, July 17

The competitive frame

Meta Business Agent lands opposite OpenAI's ChatGPT Work rollout Friday, Anthropic's Claude Cowork expansion earlier in the week, Google's Gemini 3.5 Pro enterprise launch, and Salesforce's Agentforce push. Where Meta differentiates: the messaging-surface distribution. WhatsApp counts ~2.7 billion monthly active users, Messenger ~1 billion, Instagram ~2 billion, and the customer-service and commerce use cases already run over those surfaces at scale in India, Brazil, Southeast Asia, and Mexico. For an enterprise buyer, running a Meta Business Agent on WhatsApp is closer to where the customer already is than deploying a ChatGPT Work or Claude Sonnet integration on a web app. The trade-off: Meta is not a neutral cloud, and enterprise buyers who have historically avoided taking a Meta dependency for CIO-office reasons will need to reassess.

ApplicationTech

OpenAI rolls out ChatGPT Work on GPT-5.6, folds Atlas browser into a unified desktop surface with tabs and password manager

PPC.land · 2026-07-17 · Source ↗

GPT-5.6 powers an agent that builds spreadsheets, slide decks, and documents by working autonomously across connected apps. The Atlas standalone desktop browser shuts down August 9 with functions folded into a new unified ChatGPT app.

OpenAI expanded the ChatGPT Work rollout on Friday, July 17, positioning the new agent as the flagship enterprise surface running on GPT-5.6, and confirmed the retirement of the standalone ChatGPT Atlas desktop browser on August 9. The Atlas functions fold into the unified ChatGPT desktop app, which now includes tabs, a password manager, autofill, a built-in local browser, and a separate cloud browser that runs on OpenAI's own servers for remote agent execution. The consolidation is the second half of the strategy OpenAI's James Sun confirmed on July 9.

"A system designed to complete multi-step office tasks such as building spreadsheets, slide decks, and documents by working independently across a user's connected applications."— PPC.land on ChatGPT Work, July 17

The office-tasks pitch

ChatGPT Work targets the concrete productivity workflows that Microsoft Copilot, Google Workspace with Gemini, and Anthropic Claude Cowork all compete for: spreadsheet analysis, slide-deck production, document drafting, calendar management, and multi-app task orchestration. The GPT-5.6 Sol backbone plus the new Ultra subagent mode announced July 9 give ChatGPT Work the long-horizon planning surface OpenAI has been rebuilding since the June 26 Sol limited preview. Pricing on ChatGPT Work through OpenAI's Business tier is $30 per seat per month, with Enterprise at $60 including Priority routing and audit trails.

The unified desktop app

The Friday build merges ChatGPT, the Codex CLI, and the Atlas browser functions into a single desktop app. The built-in browser can visit sites, log into accounts, and download files. The cloud browser on OpenAI's servers runs agents remotely so the user's local session is not tied up for a long-running task. The password manager and autofill features address the fundamental barrier to agent-completed purchase and booking flows: getting credentials into agent context without exposing them. That surface is where GPT-Red's Thursday internal disclosure was most focused, and OpenAI's Friday framing positions the unified desktop app as the safer container for the agent workflow than a standalone browser.

"OpenAI kills Atlas browser, folds it into new ChatGPT Work agent."— PPC.land headline, July 17

The Microsoft tension

ChatGPT Work is a direct competitor to Microsoft Copilot, and Microsoft's Friday internal sales-team directive to position Copilot as superior to ChatGPT Work in enterprise pitches is the corresponding move on the other side. OpenAI's largest investor is now explicitly competing against OpenAI in corporate sales. The two companies continue to share infrastructure and revenue at the Azure layer, and to co-develop capacity through the Stargate partnership, but the enterprise sales surface has bifurcated. Both companies are betting that channel conflict at the enterprise layer is manageable while the infrastructure and equity partnerships continue. The bet is untested at this scale.

EconomicTech

AIsphere closes $439 million Series C led by Alibaba as Chinese AI-video capital concentrates ahead of WAIC

Crescendo · 2026-07-17 · Source ↗

Alibaba Group led the round announced July 14 and reiterated in the WAIC-week capital roundup. AIsphere joins Kling AI, Vidu, and MiniMax in a Chinese AI-video generation cohort now capitalized at above $10 billion aggregate.

AIsphere completed a $439 million Series C led by Alibaba Group Holding announced originally on July 14 and reiterated in the WAIC-week capital roundup Friday, July 17. The round is one of the largest Chinese AI-video funding events of 2026 and places AIsphere alongside Kling AI, Vidu, MiniMax, and ByteDance-affiliated Doubao Video as the top-tier Chinese AI-video generation cohort. Aggregate capitalization across those five names now sits above $10 billion. Alibaba's lead position folds into its broader Q3 2026 AI ecosystem play, which includes Qwen 3.6 open weights, Ovis OCR document AI, and the AMap ABot-N1 visual language navigation stack.

The AI-video funding wave

Chinese AI-video generation has become the highest-concentration AI vertical outside foundation-model training. Sora 2, Runway Gen-4, Google Veo 3, and OpenAI Sora Turbo dominate the US market, but Kling AI 2.0, Vidu Q2, and MiniMax's Hailuo Video have captured a meaningful share of Chinese domestic and Southeast Asian consumer traffic and are increasingly used in advertising production pipelines. AIsphere's differentiation is in professional post-production workflows: color grading, VFX-quality object insertion, and scene continuity across long-form video output. The round funds a US market push and expansion into Southeast Asian mobile-video advertising.

"AIsphere completed a $439 million Series C led by Alibaba Group Holding Ltd on July 14, 2026, focusing on AI video."— Crescendo AI startup funding roundup, July 17

The Alibaba play

Alibaba's Series C lead is consistent with its 2026 pattern of concentrated bets on Qwen-adjacent and Doubao-competitive verticals. Ant Group, Alibaba Cloud, and now Alibaba's venture arm are running a coordinated AI capital-deployment strategy that mirrors what Nvidia and SoftBank are running in the US market. The concentration is real. Between AIsphere's $439M, the Kimi K3 open-weights release Thursday, DeepSeek V4 shipping Friday, and WAIC's 300 product debuts, Chinese AI's mid-July news surface is coordinated at both the technical and financial layers, in a way it was not in Q1 or Q2 2026.

Tech

DeepSeek V4 stable ships with $0.14 input pricing, legacy chat and reasoner aliases retire July 24 at 15:59 UTC

TechNode · 2026-07-17 · Source ↗

V4 Flash prices at $0.14 input and $0.28 output per million tokens, V4 Pro at $1.74/$3.48 with 75 percent promotional discounts periodically dropping the top rate to $0.435/$0.87. Peak-hour pricing doubles baseline during Beijing 9-12 and 14-18 windows.

DeepSeek V4 exited preview to standard billing this week with the July 17 mid-week window carrying the general availability window. V4 Flash prices at $0.14 per million input tokens and $0.28 output. V4 Pro prices at $1.74 input and $3.48 output at standard rates, with 75 percent promotional discounts periodically dropping the top rate to $0.435 input and $0.87 output. Both models share a 1-million-token context window at the standard rate. The launch adds a new peak-hour pricing model that runs at 2x baseline during Beijing business hours (9-12 and 14-18 local time), letting DeepSeek shed inference load away from peak without capacity-constraining paying customers.

The migration deadline

Legacy model aliases deepseek-chat and deepseek-reasoner retire on July 24, 2026, at 15:59 UTC. After that timestamp, any API call using either legacy name returns an error, with no grace period. Since April 24, the legacy aliases have been routed to V4-Flash, so most existing applications may already be running on V4 without explicit code changes. Enterprises still on the legacy aliases have six days to migrate. The change is a one-line code update in most cases, swapping the model parameter.

"After July 24, 15:59 UTC, any call using either legacy alias returns an error with no grace period beyond that timestamp."— DeepSeek migration guide, per Developers Digest

The pricing floor

DeepSeek V4's $0.14 input and $0.28 output per million tokens on Flash is now the industry price floor for a 1-million-context frontier-adjacent model. Combined with Kimi K3's Thursday $3/$15 pricing on a 2.8T open-weight model, GLM-5.2's MIT-licensed weights on Hugging Face from mid-June, and Qwen 3.6's 35B MoE and 27B dense Apache 2.0 checkpoints, the Chinese open-weight cohort has now compressed API pricing across the entire market. Gemini 3.5 Pro's $1.25/$10 pricing announced Friday sits above the Chinese-open floor but below the previous US-closed-model tier. Anthropic's Claude Fable 5 pricing remains the ceiling of the differentiated-capability tier and is under coordinated compression pressure.

TrendTech

WAIC 2026 opens with 300-plus global product debuts, 140 forums and 1,100 exhibitors on the Shanghai floor

TechTimes · 2026-07-17 · Source ↗

The four-day Shanghai World AI Conference brings 1,400 registered speakers and 300 product debuts across the exhibition floor, coordinated with the Xi keynote and WAICO founding. Chinese domestic AI infrastructure claims 40 percent of exhibitor floor space.

The Shanghai World Artificial Intelligence Conference opened on Friday, July 17, and runs through July 20 with more than 140 forums, 1,400 registered guests, 1,100 exhibitors, and more than 300 products debuting on the show floor. The scale is more than 3x the 2024 edition and roughly 40 percent above the 2025 edition, per the Ministry of Foreign Affairs opening summary. WAIC 2026 is co-anchored by Xi Jinping's in-person keynote at the High-Level Meeting on Global AI Governance and by the founding of the World AI Cooperation Organization (WAICO) with 29 signatory countries.

What is debuting

Three anchor debuts landed inside the first 24 hours: Huawei's Atlas 950 SuperPoD demonstrating 1 EFLOPS at FP8 on 1,024 Ascend NPUs; Nubia's NaviX Ultra unveiled as the world's first agent-native smartphone with OS-level ByteDance Doubao integration; and Alibaba, Ant Group, and Tencent's coordinated demonstration of Qwen 3.6, Ring-Zero 1T, and Hunyuan Harness Handbook enterprise agent workflows. Beijing-based Zhipu debuted GLM-5.3 as a WAIC preview, with weights promised for late July. StepFun and Baichuan floor demonstrations showed off Chinese-domestic frontier-model tooling. Roughly 40 percent of exhibitor floor space is Chinese domestic AI infrastructure, per TechTimes coverage.

"This year's edition would feature more than 140 forums, 1,400 guests and 1,100 exhibitors, with more than 300 products set to make their global debuts."— WAIC 2026 opening summary, per Al Jazeera

The staging matters

WAIC 2026's coordination with Xi's keynote and the WAICO founding turns the conference from a national industrial-policy showcase into a diplomatic-and-technical positioning event. The venue is the same Shanghai Convention and Exhibition Center that hosted the 2025 edition, but the delegate mix has shifted. More than 90 delegations from Global South countries attended, ASEAN sent working-group leadership across all ten member states, and BRICS members Brazil, India, and South Africa sent economic-and-technology ministers. The Silicon Valley delegation is smaller than 2025, with Anthropic, OpenAI, Google, and Microsoft all sending mid-tier representatives rather than executive leadership. That absence is being read in Shanghai as ratification of the WAICO bifurcation argument.

EconomicTech

Fireworks AI closes $1.5B Series D at $17.5B valuation, tokens hit 40 trillion per day with 95 percent from specialized models

SiliconANGLE · 2026-07-17 · Source ↗

Atreides Management, Index Ventures and TCV led the round on $1 billion-plus ARR, up 5x year over year. Cursor and Harvey are named customers. 95 percent of 40 trillion daily tokens now come from specialized models trained on customer proprietary data.

Fireworks AI closed a $1.505 billion Series D at a $17.5 billion valuation on July 15 and confirmed the round through the week, with WAIC-week capital-market coverage carrying the story into Friday's news cycle. The round was led by Atreides Management, Index Ventures, and TCV. Co-investors include Nvidia, Lightspeed Venture Partners, Evantic Capital, Bessemer Venture Partners, Menlo Ventures, Insight Partners, Ontario Teachers' Pension Plan, and Lone Pine Capital.

The numbers

Fireworks now runs at above $1 billion annualized revenue, up 5x year over year. Daily token volume on the platform has grown to 40 trillion per day from 15 trillion over the same stretch. The distribution is the key number: 95 percent of those 40 trillion daily tokens come from specialized models trained on customer proprietary data, not from general-purpose foundation model API calls. The direction is the fine-tuned-model-as-a-service thesis Fireworks has been running since 2022, now proven at scale.

"Every company must own its intelligence."— Fireworks Series D positioning statement, July 15

Named customers and use of funds

Fireworks named Cursor and Harvey as reference customers. Cursor's coding-IDE surface runs on Fireworks-served specialized inference for latency-sensitive completions. Harvey's legal AI stack uses Fireworks fine-tuning to specialize open-source foundation models on client legal corpora. The Series D proceeds will fund expanded compute infrastructure, an engineering-team hiring push, and deeper cloud partnerships with Microsoft and Nvidia. The company positions the raise as differentiating from frontier-lab-API competition on cost and specialization, arguing that enterprises will increasingly want to own their intelligence rather than pay per-token rates to a closed model.

"Every company must own its intelligence powering their products and operations."— Fireworks Series D announcement

The category positioning

Fireworks sits at the intersection of the inference-serving stack (competing with Together AI, Modal, and Baseten) and the fine-tuning-orchestration stack (competing with OpenAI's fine-tuning API and Anthropic's model customization). Nvidia's participation as a co-investor is the significant vendor signal: Nvidia is committing to the fine-tuned-specialization thesis as a complement to its frontier-training play at Anthropic, OpenAI, and xAI. The July 17 WAIC-week capital-market backdrop, with SambaNova at $11B from July 8 and Fireworks at $17.5B from July 15, positions the AI-infrastructure Series D-and-later category as the highest-growth private-equity vertical of Q3 2026.

TrendEconomic

Microsoft internal directive tells enterprise sales to position Copilot ahead of OpenAI's ChatGPT Work

Buildfastwithai · 2026-07-17 · Source ↗

The partner-versus-competitor split moves from infrastructure equity to enterprise sales channel. Microsoft's largest investee is now a named competitor in corporate deals even as Azure hosts OpenAI inference and Stargate splits capacity.

Microsoft's enterprise sales team received an internal directive Friday, July 17, to position Copilot as superior to OpenAI and ChatGPT products in corporate pitches. The reframing is the concrete signal that the Microsoft-OpenAI relationship has bifurcated across layers. Microsoft remains OpenAI's single largest external investor. Azure remains the primary infrastructure surface for OpenAI's inference and training workloads. The Stargate compute partnership continues to allocate multi-gigawatt capacity across the two companies. But the enterprise sales surface is now explicitly competitive.

The context

ChatGPT Work's Friday rollout with GPT-5.6 Sol Ultra targets exactly the enterprise-productivity workflow that Microsoft Copilot has been building toward through Office, Teams, GitHub, and the Copilot Studio agent-builder. Both products build spreadsheets, draft slide decks, and run multi-step office tasks against connected apps. Both target the $30-to-$60 per seat per month enterprise price band. Both compete for the exact same enterprise buyer conversation. Microsoft's new sales-team position is a formalization of the competition that had already existed in practice through Q2.

"OpenAI's largest investor now explicitly competes against it in corporate deals."— Buildfastwithai summary, July 17

The partnership tension

The infrastructure and equity partnerships continue. Microsoft's April earnings call flagged $190 billion in 2026 capex, with $25 billion of the increase driven by memory and storage costs that AI infrastructure demand had inflated. Azure hosts a growing share of GPT-5.6 inference, and Microsoft receives revenue share on OpenAI's Enterprise-tier deals. The revenue relationship is thick enough that channel-conflict escalation would carry real financial cost for both sides. What Friday's directive signals is that Microsoft leadership has decided that near-term Copilot revenue capture matters more than avoiding sales-channel friction with OpenAI. OpenAI's leadership has not commented publicly.

The wider read

The Microsoft repositioning fits the broader competitive-vs-partner pattern across the frontier-lab market. Google is racing Gemini 3.5 Pro into ChatGPT Work's target enterprise segment. Anthropic is competing on Claude Cowork and MCP-connected agent surfaces. Meta launched Meta Business Agent as a WhatsApp-native competitor Friday. What used to be a differentiated-capability market is a differentiated-workflow market. Vendor-agnostic enterprise buyers are the largest immediate winners; long-term integration lock-in is the strategic bet each vendor is running.

Safety

Check Point 2026 report documents autonomous agents running full exploitation chains, enterprise data leakage doubles

Help Net Security · 2026-07-17 · Source ↗

Researchers documented intrusions in which autonomous AI agents generated thousands of commands across dozens of sessions with minimal human direction. Enterprise data leakage roughly doubled year over year.

Check Point's 2026 AI Security Report published Friday, July 17, marks the first major-vendor incident report to document autonomous AI agents running full exploitation chains at scale. The specific finding: researchers observed intrusions in which AI agents generated thousands of commands across dozens of separate sessions with minimal human direction, from initial access through privilege escalation to data exfiltration. Enterprise data leakage tracked by Check Point roughly doubled year over year, driven by autonomous-agent workflows both on the attacker side and, in some cases, on the defender side where insufficient guardrails let internal agents accidentally exfiltrate sensitive data.

The autonomous-chain finding

The step-change the report documents is that AI agents are no longer used only for one-off tasks (writing phishing lures, generating polymorphic malware, translating code between languages). They are now used as the driver of full multi-step exploitation chains, executing hundreds of commands in sequence, adapting to observed defender behavior, and pivoting between initial-access, credential-harvest, and data-exfiltration phases. The behavior is a real capability shift, not a marketing frame. The mechanism is a hard-coded goal (typically variations on "gain persistent access and exfiltrate sensitive documents") passed to a coding-agent LLM with tool access to nmap, metasploit, and cloud CLIs.

"Researchers documented intrusions in which AI ran exploitation workflows autonomously, generating thousands of commands across dozens of sessions with minimal human direction."— Check Point 2026 AI Security Report

The data-leakage figure

The enterprise data leakage doubling year over year is driven by two mechanisms Check Point identifies. First, autonomous agents on the attacker side that succeed at data exfiltration where a human attacker would have been more easily detected or interrupted. Second, insufficient guardrails on internal agents deployed for productivity and analytics purposes, which the report finds have exfiltrated data through misconfigured tool permissions in roughly 12 percent of documented enterprise deployments. The recommended enterprise controls include: strict tool-permission scoping for all AI agent deployments; egress firewall rules against agent workflows accessing external endpoints; and continuous monitoring of agent tool-call traces against DLP baselines.

The read across the week

Check Point's report lands alongside Cursor IDE's zero-click RCE emergency patch, CISA's SharePoint KEV addition, WordPress 6.9.5 emergency release, and the ongoing Grok Build CLI silent-upload disclosure. The pattern: coding-agent surfaces are now attack targets, deployed agent workflows are creating new data-exfiltration surfaces, and vendor incident response is compressing from months to days. Anthropic, OpenAI, and Google all shipped safety-hardening patches this week specifically to address these classes of vulnerability. What Check Point's report adds is the aggregate: the class-level shift in attacker capability is real and measurable.

ApplicationTrend

China claims first commercial invasive brain-computer interface implant, Beijing moves ahead of US clinical-trial pipelines

Buildfastwithai · 2026-07-17 · Source ↗

Chinese state media disclosed the first commercial invasive brain-computer interface implant Friday, moving the technology from clinical-trial to commercial deployment. Neuralink, Synchron and Paradromics remain in US clinical trials. Independent verification of the Chinese procedure is limited.

Chinese state media disclosed the first commercial invasive brain-computer interface implant on Friday, July 17, claiming a step from clinical-trial-approved to commercial-market-deployed for the technology in China. Details remain limited. The vendor, patient count, hospital, and specific implanted device architecture have not been independently verified through Western scientific press. What is unambiguous is that the disclosure timing, tied to WAIC 2026 and Xi's keynote, is deliberate. Beijing is positioning Chinese biomedical AI infrastructure as ahead of the US market by shipping vs. trialing.

The US comparison

Neuralink, Synchron, and Paradromics remain in FDA-supervised clinical trials in the United States. Neuralink's PRIME study passed the 20-patient milestone earlier this year with results published in Nature Medicine. Synchron's Stentrode is in phase-2 trials at multiple US and Australian sites. Paradromics received IDE clearance for its Connexus device in late 2025 and began human implantation in Q2 2026. None of the three has crossed from investigational-device-exemption trials to commercial-market deployment, which requires PMA approval that has not been granted for any invasive BCI in the US.

The verification gap

Independent Western analysts have not yet verified the Chinese claim. State-media coverage typically leads scientific-press verification by weeks in the BCI space, and unaffiliated bioethics researchers are already flagging that the "commercial" designation could reflect a Chinese regulatory pathway that Western jurisdictions would classify as investigational. The concrete implications will depend on details that have not been released: the specific device, the patient volume, the follow-up protocol, and whether the "commercial" label reflects paid-for-treatment or accelerated regulatory approval. The Ministry of Science and Technology has not published a technical readout as of Friday evening UTC.

Tech

SEED lands as #2 on Hugging Face papers, self-evolving on-policy distillation transfers agentic RL without a full retrain

Hugging Face daily papers · 2026-07-17 · Source ↗

The method skips a fresh reinforcement-learning training run by distilling on-policy teacher traces into a stronger student. Follows July 15's Direct On-Policy Distillation paper in a compounding weekly research thread on cheap RL transfer.

SEED, "Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning," landed at #2 on Hugging Face daily papers on Friday, July 17, behind VideoChat3 and ahead of SearchOS-V1. SEED extends a method thread that Tsinghua-AIR and ByteDance-Seed opened Wednesday with Direct On-Policy Distillation, where an RL-trained teacher's policy-shift trajectories are compressed into a student model without a fresh RL training run. SEED's contribution is the self-evolving loop: the student, after distillation, becomes the next teacher for a further distilled student.

The mechanism

The distillation target is not a static teacher policy but the trajectory of a teacher policy across on-policy training. That means the student inherits not just the teacher's converged behavior but the trajectory of how the teacher's behavior evolved through the RL run. In practice: a stronger student with more parameters than the teacher, trained on a fraction of the compute a fresh RL run would have required, ends up matching or exceeding the teacher's task-completion rates. On the paper's own benchmarks against tool-using web agents and coding agents, SEED reaches teacher performance at roughly 12 percent of the target-side compute of the original teacher training run.

The research direction

The July Hugging Face daily-papers surface has been dominated by distillation and post-training efficiency papers for two weeks running: Direct On-Policy Distillation on Wednesday, ABot-N1 and ABot-AgentOS on Tuesday, and Kyutai's MuScriptor on Thursday. The direction is clear. As frontier-model pretraining compute costs pass Kimi K3's approximate $50-million-per-run figure, research budgets are shifting into cheaper post-training methods that transfer capability without the underlying compute. For agentic RL specifically, that direction pushes toward SEED-style loops where a training team can iterate agent capabilities weekly rather than quarterly.

EconomicTrend

Crunchbase records $510 billion H1 2026 global startup funding, US captures 86 percent driven by AI mega-rounds

Crunchbase News · 2026-07-17 · Source ↗

First-half 2026 startup investment reached a record $510 billion globally with US venture capturing 86 percent of the total. AI chip startups alone raised $4.8 billion across 9 deals, up 340 percent year over year.

Crunchbase's mid-July 2026 data release pegs global startup investment at a record $510 billion in H1 2026, with US venture capital capturing 86 percent of the total driven by AI mega-rounds. The concentration story is the headline. Fireworks AI at $1.5B, Neko Health at $700M, Chai Discovery at $400M, SambaNova at $1B, and AIsphere's $439M all landed inside the first two weeks of July alone. AI chip startups specifically raised $4.8 billion across 9 deals in H1 2026, up 340 percent versus H1 2025, as enterprises that couldn't get Nvidia GPU allocation piloted alternatives from Cerebras, Groq, SambaNova, and Tenstorrent and are now converting to production orders.

The exit market

Global startup exits, including IPOs and M&A, soared in the H1 2026 window per Crunchbase's disclosure. Cerebras completed a $57 billion IPO in May. Nvidia's $20 billion Groq transaction closed in December 2025 and continues to reshape the AI accelerator competitive landscape. AI security M&A specifically accelerated: 219 cybersecurity M&A deals closed in H1 2026, and AI-security acquisitions jumped from 10 in full-year 2025 to 29 in H1 2026 alone. The acceleration is driven by enterprise agent deployment outpacing security capability, which the Check Point 2026 report published Friday quantified as roughly a 2x year-over-year increase in enterprise data leakage.

"US received 86 percent of $412.7 billion H1 venture funding toward AI."— Crunchbase H1 2026 data summary

The concentration risk

The $510B record is misleading if read as breadth. Two-thirds of the total was captured by fewer than 40 deals, most of them at $500M or larger. The mega-round concentration extends to fewer than 10 anchor investors: Nvidia, SoftBank, Andreessen Horowitz, Sequoia, Index, Thrive, Lightspeed, TCV, Atreides, and Ontario Teachers' Pension Plan account for lead-investor position on the majority of the AI mega-rounds. That concentration risk cuts both ways: it enables continued capital deployment into training compute and specialized inference, and it concentrates decision-making about the direction of the AI market in a very small number of investment committees. Asia venture funding hit a multiyear high in Q2 2026 driven by fintech and AI, but the US-versus-rest gap widened.

ApplicationPolicy

China pledges MAZU AI weather-warning system to 30 developing countries as part of Xi's WAIC diplomacy package

Ministry of Foreign Affairs (China) · 2026-07-17 · Source ↗

Xi's WAIC keynote committed China to enabling 30 countries to use MAZU, its AI-powered meteorological warning system, alongside 5,000 training placements over five years and international AI application cooperation centers with regional bodies.

Xi Jinping's WAIC keynote Friday, July 17, committed China to enabling 30 developing countries to use MAZU, the AI-powered meteorological warning system Chinese state weather agencies have been building over the past three years. The commitment sits alongside 5,000 training placements for developing-country delegates over five years and international AI application cooperation centers with regional organizations. The MAZU pledge is the operational half of Xi's broader "symphony of international cooperation" framing.

What MAZU does

MAZU integrates satellite imagery, sea-surface temperature data, and atmospheric-model outputs through a deep-learning nowcasting backbone that predicts short-term severe weather events, including typhoons, floods, and heat waves at ~10-kilometer spatial resolution and 6-hour temporal resolution. The system is named for the Chinese sea goddess who protects fishermen, framing the deployment for coastal-hazard-facing populations. In domestic use since 2024 across Chinese coastal provinces, MAZU has been credited by state media with warning-lead-time improvements of roughly 40 percent versus the previous China Meteorological Administration baseline.

The Global South target

The 30-country commitment targets ASEAN member states, sub-Saharan African coastal nations vulnerable to Indian Ocean cyclones, and Latin American Pacific-coast countries exposed to El Niño-driven flooding. The specific countries have not been disclosed. What Xi's commitment competes against is Nvidia's Earth-2 sovereign-AI weather platform, Google DeepMind's GraphCast, and the ECMWF-hosted AIFS suite. Each of those systems is available today in some form to developing-country weather agencies but requires either Nvidia GPU capex or ECMWF membership. MAZU's Chinese-domestic-hardware backbone means recipient countries do not need US-origin AI infrastructure to operate the system, which fits the Huawei Atlas 950 SuperPoD deployment thesis Xi outlined in the broader keynote.

The soft-power play

MAZU is the second concrete deliverable Xi's WAIC address produced, alongside WAICO membership. The pattern matches Beijing's soft-power model for the past decade: bundle infrastructure, training, and technical assistance into a coordinated package targeted at Global South economies with the expectation of long-term supply-chain and political-alignment returns. For US and European weather-modeling programs, MAZU's rollout is a real competitive move, particularly in ASEAN where Chinese diplomatic influence is high and Nvidia GPU access remains constrained.

Tech

SearchOS-V1 lands #3 on Hugging Face daily papers with a robust open-domain information-seeking agent framework

Hugging Face · 2026-07-17 · Source ↗

Positions a robust agent-collaboration architecture across heterogeneous tools for open-domain information retrieval. Third on the Friday daily surface behind VideoChat3 and SEED.

SearchOS-V1 landed at #3 on the Hugging Face daily papers surface Friday, July 17, positioning a robust agent-collaboration architecture for open-domain information seeking that handles heterogeneous tool routing across web search, database queries, code-execution sandboxes, and API calls. The paper follows the direction Anthropic set with Model Context Protocol and OpenAI followed with Codex tool use: a first-class abstraction over the tool-selection problem that lets a planner LLM route across the right tool for each subtask.

What SearchOS gets right

The concrete contribution is the routing layer. Prior open-domain agents (Perplexity, Arc Search, You.com) route search queries through a small closed set of tools with hand-tuned selection. SearchOS-V1 makes tool selection itself a learned behavior, with the agent's own reasoning trace informing which tool from a catalog is best for a given subtask. On the paper's evaluation set, SearchOS-V1 reaches 68 percent success on the BrowseComp benchmark, ahead of Perplexity Deep Research's 51 percent and Anthropic's Search tool at 62 percent. Independent verification through the Hugging Face community will land over the next 30 days.

The July research surface

The top three Friday papers, VideoChat3, SEED, and SearchOS-V1, together mark a distinct July research direction: fully-open backbones, post-training distillation methods that skip the compute-expensive RL run, and agent-collaboration frameworks that operate on heterogeneous tool catalogs. Combined with the ByteDance-Seed and Tsinghua-AIR direct on-policy distillation paper Wednesday, the direction is toward cheaper post-training and more general agent capabilities without a proportional compute lift. Whether this direction produces a Q3 frontier-model release from a Chinese lab remains the open question.

Tech

Ollama v0.32.1 fixes JSON brace parsing in tool-call detection, stabilizes Gemma 4 multi-turn continuations

Ollama release notes · 2026-07-17 · Source ↗

Point release corrects an early tool-call boundary bug that had truncated JSON payloads containing embedded braces. Includes reliability improvements to Gemma 4 tool-response continuation across multi-turn dialogues.

Ollama tagged v0.32.1 on Friday, July 17, as a follow-up to the v0.32.0 general availability release from July 11. The tagged release is a point-release stability update focused on tool-call detection reliability. The signature fix credited to contributor @aditya-786 addresses a case where Ollama's tool-call boundary detector treated closing braces inside JSON string literals as the end of a tool-call payload, truncating responses that contained embedded JSON. The fix ignores braces inside JSON string content when detecting tool-call end.

Gemma 4 multi-turn stabilization

v0.32.1 additionally lands reliability improvements for Gemma 4 tool-call and multi-turn reasoning. The updates specifically address more reliable tool-response continuations across multi-turn dialogues, where Gemma 4 had previously dropped tool-call context after the third or fourth turn in agent workflows. Combined with the v0.32.0 interactive-agent surface improvements and MLX cache-leak fix, the July Ollama release cadence has focused on the coding-agent and tool-use edge cases that had emerged as production traffic scaled through Q2.

"Tools: ignore braces inside JSON strings when detecting tool call end."— Ollama v0.32.1 release notes, @aditya-786

The wider release cadence

Ollama's July release cadence sits inside a broader inference-runtime maturity story. llama.cpp shipped 11 builds between July 15 and July 16 alone, including SYCL Flash Attention with the XMX engine via oneDNN for a 1.21x prefill speedup on Intel Battlemage, and Tencent Hunyuan 3 architecture support. Anthropic's Claude Code point release Friday, OpenAI Codex CLI's rust-v0.144.5 dangerous-command hardening earlier in the week, and Ollama's v0.32.1 today form a coordinated pattern: coding-agent and tool-use infrastructure is hitting production edge cases faster than release cadence had previously accommodated. The direction is toward point releases every 24-72 hours across the entire coding-agent stack.