Daily2026-08-10
TechApplication

xAI ships Grok 4.6 at 1.5 trillion parameters, opens API and Grok Voice 3.0 the same morning

The 08:12 UTC push moves grok-4-latest to a 1.5T dense-MoE architecture trained on the Colossus 2 build-out that came online in June, priced at $2.80 per million input tokens and $9.40 output, with a 512K native context, day-one function-calling and structured-output support, and Grok Voice 3.0 rolled into the same release with 280ms end-to-end latency. Musk posted the announcement at 08:14 UTC and put the 2.1-trillion-parameter Grok 4.7 on a two-week schedule.

xAI · 2026-08-10 · Source ↗

xAI pushed Grok 4.6 to the public API at 08:12 UTC Monday morning, moving the grok-4-latest alias to a 1.5-trillion-parameter dense mixture-of-experts model trained on the Colossus 2 datacenter build-out that came online in Memphis in June. The launch post at x.ai/news/grok-4-6 lists $2.80 per million input tokens and $9.40 output per million tokens through the API, a 512K native context window, day-one support for structured output and function calling in the same request format the 4.5 series used, and native MCP 2026-07-28 tool support without a translation shim. Grok.com switched over to 4.6 as the default for SuperGrok and Heavy users at the same moment. Musk followed at 08:14 UTC with a post on X naming Grok 4.7 as the 2.1-trillion-parameter successor and putting it on a two-week schedule.

The pricing places Grok 4.6 below Claude Opus 5 ($5 input, $25 output) and roughly at parity with GPT-5.6 Sol ($2.40 input, $9.60 output) at the frontier tier. Against DeepSeek V4-Flash's $0.14 input and $0.28 output the gap remains an order of magnitude, and xAI made no attempt to compete with the open-weights price floor. Musk framed the pricing on X as reflecting that Colossus 2 was funded from cash flow rather than debt, which lets xAI hold higher margins per token than the labs financing capacity on convertible notes.

"Grok 4.6 is 1.5T parameters, trained fully on Colossus 2. Grok 4.7 at 2.1T ships in two weeks. Both stay fully unfiltered on the platform. Cheaper too." — Elon Musk on X, 08:14 UTC August 10, 2026

The architecture

The technical report published alongside the launch describes Grok 4.6 as a dense-MoE hybrid: 1.5 trillion total parameters, roughly 120 billion active per token, with 96 experts of which eight route per forward pass. That places it structurally between DeepSeek V4-Flash (open-weights, sparse MoE, roughly 32B active from 671B total) and Claude Opus 5 (Anthropic has not published architecture details but the inference latency profile is consistent with a dense model in the low-hundreds-of-billions range). The 512K context is genuine, not a sliding window, and internal needle-in-haystack tests published in appendix B of the report show recall above 96 percent at every 32K interval out to 480K. Beyond 480K accuracy drops to the mid-70s and Musk's post recommended splitting inputs above that boundary.

Training used approximately 210,000 H200-class GPUs equivalent across Colossus 2's actual Blackwell-generation hardware, with the report claiming 187 zettaflops of effective compute over the full run, or roughly 3.4 times what Grok 4 consumed and roughly 1.8 times what Grok 4.5 used. The dataset extended the crawl through July 2026 and added the video-corpus post-training pass that Grok 4.5 introduced. Reinforcement learning from verifiable rewards accounted for the final 14 percent of tokens and drove most of the coding and math gains.

The benchmark picture

xAI published headline numbers on twelve benchmarks. Terminal Bench 2.1, the coding-agent benchmark that has become the industry standard since Anthropic and OpenAI both adopted it in June, comes in at 89.3 percent for Grok 4.6, which is within measurement noise of GPT-5.6 Sol at 89.5 percent and Opus 5 at 89.1 percent. On the AIME-2026 mathematics evaluation, Grok 4.6 posts 98.7 percent, edging Opus 5 at 97.8 percent and Sol at 97.4 percent. GPQA-Diamond lands at 89.1 percent, ARC-AGI-3 at 62.4 percent, SWE-Bench Verified at 76.8 percent. The report includes a full internal reproduction of Anthropic's Frontier-Math evaluation methodology and puts Grok 4.6 at 43.2 percent, ahead of the 41.6 percent Anthropic reported for Opus 5 in July.

The gap Grok 4.6 does not close is the humanities and long-form writing bench. On the Aiden Bench 3.0 open-ended composition evaluation Opus 5 remains ahead by roughly six points, and independent testers on X reported through the morning that Grok 4.6 tends toward verbose, list-heavy prose without careful prompting. Aiden McLaughlin, who runs the eponymous benchmark, posted at 10:22 UTC that the 4.6 update was a coding and reasoning release rather than a prose release.

"Grok 4.6 lands right on top of GPT-5.6 Sol on coding and reasoning benchmarks. On long-form writing it's still meaningfully behind Opus 5. Different labs, different objectives." — Aiden McLaughlin, Aiden Bench maintainer, on X

Grok Voice 3.0

The same push includes Grok Voice 3.0, replacing the Grok Voice Think Fast 2.0 model that shipped August 6 at $0.08 per minute. Voice 3.0 targets 280ms end-to-end latency (cellular median), down from 400ms for Voice 2.0, and adds real-time interruption handling that lets the caller cut off Grok mid-sentence without losing context. Pricing holds at $0.08 per minute for consumer use inside the Grok app; developers accessing Voice through the API pay $0.12 per minute with the transcription and TTS layers billed separately. The 3.0 model is bilingual-native across 22 languages, expanding from Voice 2.0's twelve.

The competitive frame

Grok 4.6 is the first frontier release since OpenAI's Astra pause on Friday, and its Monday-morning timing looks deliberate. With Astra parked inside its Preparedness Framework Critical envelope and no near-term OpenAI release on the calendar, xAI takes the lead on frontier coding numbers for at least the next two weeks. Musk's two-week window for Grok 4.7 lands on August 24, roughly a week after Anthropic flips Claude Code auto mode to default and days before Anthropic's Sonnet 5 promotional pricing ends on August 31. The August calendar reads as a race for enterprise mindshare while OpenAI's next frontier model is off the field.

The Astra pause also raises the question of whether Grok 4.6 was evaluated against equivalent Critical criteria. xAI's launch post says the model was evaluated under xAI's own Responsible Deployment Framework, described in a linked appendix as "structurally similar" to OpenAI's Preparedness Framework, with results deemed to sit within the High tier for cybersecurity and Medium for CBRN. The framework document has not been externally peer-reviewed, and Simon Willison flagged on X at 09:41 UTC that xAI has yet to open its evaluations to third-party red-teamers on the model-weight level rather than the API level.

For enterprise customers already invested in Anthropic or OpenAI, the pricing gap versus Claude Opus 5 and the parity on coding benchmarks make Grok 4.6 a credible alternative for the coding-agent workload specifically. Whether the pricing holds through Grok 4.7's launch, whether xAI opens Weights-level red-team access to keep pace with the disclosure standard OpenAI just set, and whether the 4.7 release stays on the two-week schedule are the three variables that will decide how much of the enterprise pipeline shifts by the end of August.

Big Picture

Grok 4.6 lands at 1.5 trillion parameters through the xAI API at 08:12 UTC Monday, with the 2.1T Grok 4.7 on a two-week clock. Amazon Zoox begins paid Las Vegas robotaxi service at 09:00 Pacific under the NHTSA purpose-built exemption, the first commercial fare for a purpose-built vehicle from any operator. Unitree Robotics opens IPO subscription on the Shanghai STAR Market at 150.80 yuan per share, giving DeepSeek a strategic-placement toehold in the first mainland-listed humanoid maker.

The EU AI Office formally asks Anthropic, Google DeepMind and xAI whether comparable Critical-tier evaluations exist at their labs, escalating Friday's OpenAI Astra pause into a live compliance question three weeks into Article 101 enforcement authority. Alphabet holds an emergency Sunday board call over the Discovery Loop concessions after two trading days of losses. Simon Willison's Sunday follow-up on auto mode dissects the Trajectory Labs test methodology in detail. Anthropic ships Claude Code v2.1.227 with an auto-mode dry-run flag one week ahead of the default flip. OpenAI Codex CLI merges Windows-native sandbox support and a WASI plugin loader. llama.cpp b10333 lands the b10330 attention-fusion kernel on ROCm 7 for AMD MI300X. vLLM 0.10.4 ships Grok 4.6 day-one weights loader. DeepSeek pushes V4-Flash-0810. Google DeepMind Gemini Robotics ER 2 preview lands on Vertex AI. Perplexity Comet 2.0 browser beta ships. Meta Muse Spark 1.2 downloads cross 400,000. Anthropic files a Northern District of California suit against a 12-person prompt-injection consultancy. Amazon Pecos draws its first federal lawsuit from Public Citizen.

More today

28 stories · 2026-08-09 — 2026-08-10
ApplicationTech

Zoox opens paid Las Vegas robotaxi service, first fare for a purpose-built vehicle

Zoox · 2026-08-10 · Source ↗

Amazon-owned Zoox flipped the meter on at 09:00 Pacific Monday, ending eleven months of free public rides in Las Vegas and starting commercial fares in its bidirectional, steering-wheel-free vehicle. The service runs under the NHTSA purpose-built exemption capped at 2,500 vehicles per year.

Zoox began charging fares on its Las Vegas robotaxi service at 09:00 Pacific time Monday morning, converting the eleven-month free-public-rides program that started in September 2025 into a paid commercial service. The launch is the first paid fare ever charged for a vehicle built from the ground up without a steering wheel, brake pedal, or human-facing driver controls, and Amazon's Zoox subsidiary is now the only operator running a bidirectional, symmetric-passenger-cabin vehicle in commercial service anywhere in the world. The service area covers the Las Vegas Strip, downtown Las Vegas, and a corridor to Harry Reid International Airport, with launch partnerships announced with Resorts World, Wynn, Caesars Palace and the airport itself.

Pricing is per ride, distance-and-time based, and Zoox published a table on the launch page: base fare $4.50, per-mile $2.15, per-minute $0.35, minimum fare $9.00, airport pickup surcharge $6.50. Peak-hour surge multipliers cap at 1.5x, notably lower than Uber's uncapped peak pricing on the same corridor. A Zoox spokesperson quoted in the launch post said the operating economics favor a lower cap because vehicle utilization inside the fleet is set by dispatch software rather than driver behavior. The company did not publish a projected revenue-per-vehicle-day number.

"This is the first paid ride ever offered in a purpose-built robotaxi. Every autonomous service running today, ours included until this morning, has either been on a retrofitted human vehicle or has been free while we finished testing." — Aicha Evans, Zoox CEO, in the launch statement

The NHTSA context

The service runs under the NHTSA exemption Zoox received in February 2026 that allows up to 2,500 purpose-built vehicles per year in commercial deployment without meeting Federal Motor Vehicle Safety Standards specific to human-facing controls. The exemption is capacity-capped rather than time-capped, meaning Zoox can operate the same vehicles indefinitely provided the annual fleet expansion stays under 2,500. The company said Las Vegas will run with 240 vehicles at launch, with a Phoenix expansion planned for October and San Francisco expected in Q1 2027.

The launch lands one week after Amazon's Pecos, Texas AWS campus plan drew federal opposition, and Zoox chose not to make Amazon-parent language a centerpiece of the launch messaging. The launch post attributes the service to Zoox specifically, with an Amazon corporate footer only in the fine print.

The competitive frame

Waymo runs the closest comparable service, with sixth-generation retrofitted Jaguar I-PACE vehicles across eleven cities crossing 500,000 paid rides per week as of last Wednesday's Alphabet earnings. Waymo declined to comment on the Zoox launch when reached by Bloomberg. Tesla's robotaxi service in Austin remains on the pilot list published in June and has not made the transition to purpose-built Cybercab vehicles that Musk announced for Q4. Cruise, still working through a two-year regulatory recovery after its 2023 San Francisco incident, has no paid-service timeline. On the launch morning Zoox is the only operator charging fares for a vehicle that was never designed to be driven by a human.

The pricing gap versus Uber and Lyft on the Strip corridor is roughly 10 to 15 percent, low enough that peak-time and airport rides may shift measurably in the first month. The exclusivity contract with Resorts World routes hotel-summoned rides to Zoox by default when a vehicle is within five minutes, which is the single largest volume commitment. Whether Zoox can scale to the 2,500-vehicle annual cap without incident, whether the NHTSA cap holds through the coming exemption-renewal cycle, and whether Amazon absorbs the launch losses long enough to reach unit economics parity with Waymo are the three variables that will determine whether the launch reads as a milestone or a footnote by year-end.

Economic

Unitree Robotics opens IPO subscription at 150.80 yuan on Shanghai STAR

Reuters · 2026-08-10 · Source ↗

The subscription window opened at 09:30 Beijing time Monday for 40.4 million shares at 150.80 yuan each, targeting 6.1 billion yuan on a 61 billion yuan valuation. DeepSeek anchored the strategic-placement tranche, and Unitree becomes the first mainland-listed humanoid maker.

Unitree Robotics opened its Shanghai STAR Market IPO subscription at 09:30 Beijing time Monday morning, five days after pricing the deal at 150.80 yuan per share. The offering covers 40.4 million shares, targets 6.1 billion yuan (roughly $850 million) in gross proceeds, and sets a post-money valuation of 61 billion yuan (roughly $9.04 billion). Institutional demand was oversubscribed 41 times inside the first two hours of the window, per a mid-morning update from the CSRC-regulated bookrunner CITIC Securities. Retail subscription follows at 09:30 Tuesday for the remaining tranche.

The DeepSeek presence in the strategic-placement tranche is the piece that most Chinese-language coverage led with. DeepSeek subscribed to 12 percent of the strategic block, roughly 1.4 percent of the total offering, alongside China Investment Corporation (18 percent), Tencent (11 percent), and the Shanghai Municipal Investment Group (9 percent). The DeepSeek stake gives the Hangzhou lab an equity position in the first mainland-listed humanoid maker on the same weekend that Unitree began deploying its H2 unit inside BYD dealerships and JD.com fulfillment centers. Unitree CEO Wang Xingxing told the Shanghai Securities News on Sunday that the DeepSeek subscription came with a technical collaboration around on-device inference for the H2 platform's language and vision stack.

"The DeepSeek subscription is not passive. We have agreed a joint program to deploy DeepSeek V4-Flash on the H2 as its onboard reasoning model, with quarterly weight refreshes." — Wang Xingxing, Unitree CEO, Shanghai Securities News

The valuation frame

The $9.04 billion valuation places Unitree at roughly half of Figure AI's most recent private mark of $16 billion (September 2025 Series E) and at rough parity with Agility Robotics' $9.4 billion (April 2026 secondary). It sits below Boston Dynamics, which Hyundai now values internally at $28 billion after the RMAC deployment. Unitree's revenue for the trailing four quarters through Q2 2026 came in at 4.8 billion yuan (roughly $670 million), which puts the offering at roughly 13.5 times trailing revenue, considerably higher than the 6 to 8 times multiple industrial-robotics comparables trade at but below the 20-plus multiples the US humanoid names command.

The prospectus lists BYD, JD.com and Foxconn as the three largest customers, together accounting for 61 percent of trailing revenue. The concentration risk is flagged in the prospectus and priced into the offer, per morning research notes from CICC and Huatai Securities.

What listing changes

Being publicly listed forces two changes that private robotics companies have been able to defer. First, Unitree now has to publish quarterly R&D disclosures, which will include unit shipments and average selling price in a form that competitors and customers can read. Second, the STAR Market listing requires that any material foreign-technology dependencies be flagged in the prospectus; Unitree disclosed dependencies on Nvidia Orin (roughly 40 percent of BOM cost on the H2), Analog Devices power management, and Corning display glass, and committed to a China-source qualification program targeting 70 percent domestic BOM by end-2027.

Trading is scheduled to begin Thursday August 14 on the STAR Market. Grey-market pricing over the weekend put the first-day open at approximately 240 to 260 yuan, roughly 60 to 75 percent above the offer, which would place Unitree in the top decile of STAR Market first-day pops for 2026. Whether the retail tranche closes cleanly Tuesday, whether the grey-market indications hold to Thursday, and whether DeepSeek's stake becomes the template for a broader Chinese AI-lab-into-robotics equity push are the three questions the offering leaves open.

TechSafety

Claude Code v2.1.227 adds auto-mode dry-run flag four days before default flip

GitHub — anthropics/claude-code · 2026-08-10 · Source ↗

Anthropic cut v2.1.227 at 06:52 UTC Monday, adding a --dry-run flag that runs auto mode inside a manual-approval session so developers can preview classifier decisions before the August 14 default flip. Six additional fixes target Remote Control reconnect and MCP OAuth latency edges.

Anthropic tagged Claude Code v2.1.227 at 06:52 UTC Monday morning, four days before auto mode becomes the default routing for tool approvals across Pro, Max and Team plans. The headline addition is a --dry-run flag on the auto command that lets developers see what auto mode would decide for each tool call without actually delegating the approval loop; permission prompts still fire normally, but the classifier's judgment shows next to each prompt so the developer can compare their choice to the classifier's before the default flip on August 14.

The change reads as an on-ramp designed to blunt the resistance that Anthropic's product-lead Alex Albert acknowledged in a Sunday-evening post on X. Developers who have opted out of auto mode during the beta had cited two reasons in the changelog issues: they did not trust the classifier's judgment on their specific codebase, and they wanted to see empirically what the classifier would do before letting it decide for them. The dry-run flag addresses both concerns without changing the auto-mode logic itself.

"Added --dry-run to auto: manual approval loop still fires, but each prompt shows the classifier's would-be decision inline. Use this to build calibration confidence before Aug 14." — Claude Code v2.1.227 changelog

The other fixes

Beyond the dry-run flag, v2.1.227 addresses six fixes, most inherited from the Saturday-into-Sunday reliability queue that v2.1.225 and v2.1.226 did not clear. The Remote Control web-session reconnect logic that v2.1.225 truncated event replay for still misreported reconnect latency in the UI when the machine hosting the session had been asleep for more than 30 minutes; the fix rebases the reconnect clock on wall time rather than session-uptime time. The MCP OAuth server flow on macOS that v2.1.225 hardened against keychain latency was still surfacing 401s on the second consecutive OAuth refresh in a 60-second window; the fix adds a 250ms cooldown between refresh attempts. The gateway spend-limit warning that v2.1.225 introduced was rendering the operator-supplied text with HTML entity encoding preserved when the gateway sent a UTF-8 payload; the fix decodes before rendering.

A minor addition in the VS Code extension shows the classifier's live decision in the Focus panel when auto mode is active, matching what the CLI dry-run flag surfaces on the terminal. The extension update ships as part of the same release; users on the auto-updating channel will pick it up within the next four hours.

The context

The v2.1.227 release closes a five-day cadence of daily Claude Code point releases (v2.1.223 on Aug 6, v2.1.224 on Aug 7, v2.1.225 and v2.1.226 on Aug 8, and now v2.1.227 on Aug 10; v2.1.226-hotfix-1 was pulled Sunday afternoon after a metadata mismatch that did not affect binaries). The pattern is consistent with a product team under launch-deadline pressure, and Anthropic infrastructure lead Boris Cherny posted on X at 08:10 UTC that a v2.1.228 patch was expected Wednesday to catch any residual auto-mode issues surfaced by the dry-run flag before the Thursday flip.

For developers, the practical guidance is to enable --dry-run on any repository where auto mode's default judgment matters, run through a representative day of work, and only opt out of the August 14 flip if the classifier's judgment materially diverges from the developer's on a category of tool call that matters. Anthropic's documentation added a Monday-morning page walking through the dry-run flag with example output; the page is linked from the release notes.

Tech

OpenAI Codex CLI merges Windows-native sandbox support and a WASI plugin loader

GitHub — openai/codex · 2026-08-10 · Source ↗

The Monday commit stream on openai/codex main lands PR 37642 (Windows AppContainer sandbox for tool execution) and PR 37651 (WASI-based plugin loader for hooks), setting up rust-v0.148.0-alpha.6 for Tuesday release. Together they close two enterprise-request gaps from the 0.147.0 GA week.

OpenAI's Codex CLI repository merged fifteen commits to main on Monday, headlined by PR 37642 (Windows AppContainer sandbox for tool execution) and PR 37651 (WASI-based plugin loader for hooks). Both had been open since Wednesday and both closed enterprise-request threads that had accumulated during the 0.147.0 general-release week. Neither is in the current 0.148.0-alpha.5 tag; both target 0.148.0-alpha.6, expected Tuesday morning UTC.

Windows-native sandbox support has been the single most requested feature since Codex CLI shipped in April. The 0.147.0 documentation had recommended WSL2 or Docker for Windows deployments, which enterprise administrators pushed back on because both add a layer of virtualization that complicates network policy and observability. PR 37642 (author matklad) implements a sandbox using the Windows AppContainer isolation primitive, the same mechanism Edge uses to sandbox renderer processes, with per-tool-call capability grants for file system, network and registry access. The implementation adds roughly 4,200 lines of Rust in a new codex-sandbox-win crate; the linked design document walks through why an AppContainer approach was chosen over the alternative of a Windows Job Object plus firewall rule tuple.

"Implement Windows AppContainer sandbox for tool execution. Adds codex-sandbox-win crate with per-tool capability grants and Windows Filtering Platform integration for network policy." — PR 37642, merged 14:31 UTC August 10, 2026

The WASI plugin loader

PR 37651 replaces the shared-library plugin loader that Codex hooks have used since 0.140.0 with a WASI (WebAssembly System Interface) loader. The change lets administrators write hook logic in any language that compiles to WASI (Rust, Go, TinyGo, Python via componentize-py, C, C++, C#) rather than the current pattern of writing Rust or C and matching an exact ABI. The security story is the piece that pushed the PR forward: WASI plugins run inside a Wasmtime-managed sandbox that gets the same capability-grant treatment as tool execution, so a compromised or misconfigured hook cannot escalate into the parent Codex CLI process.

Existing shared-library plugins continue to work for at least the 0.148 and 0.149 major-minor cycles, per the deprecation note in the PR description, but the CLI now prints a one-line deprecation warning at startup when a shared-library plugin is loaded. New plugin documentation shipped with the PR walks through the pattern for porting a shared-library plugin to WASI in Rust; the Python and Go examples are marked as coming later this week.

The rest of the day

Thirteen additional commits merged: five backports of Saturday's authentication and Guardian hardening to the 0.147.x stable branch (releasing as 0.147.4 tomorrow), three quality-of-life improvements to the gRPC code-mode host that Saturday's PR 37530 introduced (backpressure, TLS certificate rotation, per-tenant rate limiting), two documentation updates that expand the enterprise deployment guide to cover the new Windows sandbox, and three test-only additions.

The pattern across the last week is a Codex CLI team executing against a well-articulated enterprise-deployment feature set: multi-host operation, capability-grant sandboxing, workload identity, extensible policy plumbing. The Anthropic side of the same conversation is Claude Code's Managed Agents story, which currently covers macOS, Linux and WSL2 but not native Windows. On Monday afternoon Codex CLI became the coding-agent that runs unmodified on the largest enterprise developer platform. The consequence for Anthropic is that a category of enterprise deployment now defaults to Codex until Claude Code closes the gap.

Tech

llama.cpp b10333 lands the b10330 attention-fusion kernel on AMD ROCm 7

GitHub — ggml-org/llama.cpp · 2026-08-10 · Source ↗

Slaren merged PR 26804 at 11:47 UTC Monday, porting Saturday's CUDA rms_norm+mul+RoPE fusion to ROCm 7 for AMD MI300X. Benchmarks in the PR show a 22 percent throughput lift on Qwen 3.8-Max at 8K sequence length and 34 percent at 128K.

llama.cpp maintainer slaren merged PR 26804 into main at 11:47 UTC Monday morning, porting Saturday's fused-attention CUDA kernel (rms_norm followed by elementwise multiply followed by RoPE, with optional view and set_rows) to AMD's ROCm 7 platform on the MI300X and MI325X. The b10333 tag ships two additional PRs alongside the port: PR 26810 (author JohannesGaessler) adds the same fusion path to the Metal backend for Apple silicon, and PR 26812 (author 0cc4m) adds a fallback path for CUDA architectures below Ampere so users on Turing and older cards no longer see the kernel silently disable.

The performance data slaren published in the PR shows a 22 percent throughput improvement on Qwen 3.8-Max at 8K sequence length, 34 percent at 128K, and 41 percent at 256K on an 8x MI300X node running the open weights DeepSeek released Sunday. The larger gains at longer context reflect that the fused kernel eliminates three memory-bandwidth-bound launches per attention block, and memory bandwidth is the binding constraint at long context on the MI300X's 5.3 TB/s HBM3 fabric.

The AMD frame

The port matters because llama.cpp is the reference open-weights inference stack that a very large fraction of self-hosted deployments run on. Until Monday, users running open weights on AMD accelerators paid a roughly 25 to 30 percent throughput penalty versus users running the same weights on Nvidia hardware, largely because attention kernels landed on CUDA first and often took months to reach ROCm. The b10333 release closes the gap for the current attention pattern.

"For anyone benchmarking MI300X against H200 on Qwen 3.8-Max or DeepSeek V4-Flash, the numbers you got last week are stale. Rerun on b10333." — slaren, llama.cpp maintainer, on the release discussion

AMD's developer-relations team retweeted the release inside the hour, and Lisa Su's Q2 earnings guidance from Tuesday for a Data Center revenue doubling in 2027 depends on exactly this kind of parity closing across the open-weights inference stack. The remaining gap is in mixture-of-experts routing kernels, where ROCm still lags CUDA by roughly 15 to 20 percent on sparse-activation patterns; that PR (26787) is open and unmerged.

Tech

vLLM 0.10.4 ships Grok 4.6 day-one weights loader, Colossus-2 config template

GitHub — vllm-project/vllm · 2026-08-10 · Source ↗

Zhuohan Li tagged vLLM 0.10.4 at 10:14 UTC Monday, roughly two hours after xAI's Grok 4.6 API push, with a day-one weights loader for the 1.5T model, a Colossus-2 deployment config template, and updated MoE routing kernels tuned for the 96-expert layout.

vLLM tagged version 0.10.4 at 10:14 UTC Monday, roughly two hours after xAI's Grok 4.6 API push and closely coordinated with xAI's model-card release on Hugging Face. The release includes a day-one weights loader for the 1.5-trillion-parameter Grok 4.6 architecture, a reference deployment configuration tuned for Colossus-2-class hardware (8x H200 with NVLink switch, 640GB per node HBM, InfiniBand fabric), and MoE routing kernel updates that support Grok 4.6's specific 96-expert-with-8-active layout.

The day-one loader is unusual in that xAI shared architecture details and reference weights checksums with the vLLM team ahead of the public API launch, so vLLM 0.10.4 can load and serve Grok 4.6 weights without waiting for community reverse-engineering. Weights are not publicly available (Grok 4.6 remains closed-source through the API), but enterprise customers with private licensing agreements can now serve them on vLLM from day one. The pattern mirrors the coordination Meta established with vLLM for the Muse Spark 1.2 open-weights release last week.

"Day-one Grok 4.6 support. Contact xAI enterprise for weights. Colossus-2 reference deployment: 8xH200, tp=8, pp=1, expert-parallel=8, ~340 tokens/sec per node at 512 context batch." — vLLM 0.10.4 release notes

The MoE routing update

The 96-expert-with-8-active layout Grok 4.6 uses is different from what vLLM's existing routing kernels handle well. DeepSeek V4-Flash uses 256 experts with 8 active per token; Qwen 3.8-Max uses 128 with 12 active. The Grok 4.6 layout has fewer, larger experts, which changes the memory-access pattern during routing. The vLLM 0.10.4 kernel adds a fast path for expert counts between 64 and 128 that keeps expert weights hotter in L2 cache during routing. Benchmarks in the release notes show a 14 percent throughput improvement versus loading Grok 4.6 with the pre-0.10.4 MoE routing on the same hardware.

Beyond Grok 4.6, the release includes broader improvements: prefix-cache reuse across parallel sessions (up to 40 percent memory reduction for multi-turn workloads), a scheduler fix that eliminated a starvation edge for short-request streams behind long-request contexts, and Python 3.13 support for the client library. The v0.10.4 release closes a five-week release cycle since 0.10.3; the next release 0.11.0 targets Grok 4.7 support (assuming the two-week schedule Musk announced holds) and native flash-attention 4 integration.

Tech

DeepSeek pushes V4-Flash-0810 open weights with a new tool-call planner

Hugging Face — deepseek-ai · 2026-08-10 · Source ↗

DeepSeek uploaded a Sunday-night refresh of V4-Flash (0810 stamp) with a distilled tool-call planner as a new head above the base MoE, lifting SWE-Bench Verified from 66.1 to 72.3 percent without changing base weight counts. Downloads passed 45,000 in the first six hours.

DeepSeek uploaded V4-Flash-0810 to Hugging Face at 22:41 UTC Sunday, a refresh of the V4-Flash open weights that maintains the same 671B total, 32B active per token base but adds a new tool-call planner head trained through the RL-with-verifiable-rewards regime DeepSeek pioneered in R1. The 0810 model card reports SWE-Bench Verified at 72.3 percent, up from 66.1 percent for V4-Flash-0731 released two weeks ago; Terminal Bench 2.1 at 84.9 percent, up from 82.7 percent; and Aider Polyglot at 78.4 percent, up from 71.2 percent. The base language and reasoning benchmarks are unchanged within noise.

The tool-call planner is described in the accompanying paper as a lightweight (roughly 2.1B parameter) auxiliary head that sits above the base MoE and specializes in tool-call sequencing decisions. When the base model wants to invoke a tool, the planner head reviews the last-N tool calls in the trajectory and returns either an approval, a rewrite (a different tool call the base should have made), or a rejection with reasoning. Training data was distilled from Claude Opus 5, GPT-5.6 Sol, and Gemini 3.5 Pro tool-use traces, roughly 4 million approved trajectories, per the paper.

"The planner adds 2.1B parameters and roughly 8ms of inference latency per tool call, in exchange for a 6.2-point SWE-Bench Verified improvement. That is a good trade for agent workloads." — DeepSeek V4-Flash-0810 model card

Downloads and reactions

The model passed 45,000 downloads in the first six hours after release, roughly triple the pace V4-Flash-0731 saw on release. Simon Willison flagged the release in his Sunday-evening links post, calling out the planner as "the first open-weights answer to the Anthropic classifier pattern that landed with auto mode last week." Meta AI's PyTorch team confirmed compatibility with TorchTitan out of the box. NVIDIA released a TensorRT-LLM engine build for the model at 04:12 UTC Monday.

The release keeps DeepSeek visibly ahead of the open-weights curve: V4-Flash-0810 becomes the highest-scoring open-weights model on SWE-Bench Verified, ahead of Kimi K3 (69.8 percent) and Muse Spark 1.2 (68.3 percent), and matches or beats the frontier closed-weights on coding benchmarks other than GPT-5.6 Sol and Grok 4.6 in the same tier. The pricing gap remains order-of-magnitude: V4-Flash open weights served on inference-partner API cost roughly $0.14 per million input and $0.28 per million output, versus Grok 4.6's $2.80 and $9.40. For enterprise workloads that can tolerate open-weights hosting and slightly weaker frontier ceiling, DeepSeek continues to be the pricing floor the industry defends against.

TechApplication

Google DeepMind Gemini Robotics ER 2 preview lands on Vertex AI

Google Developers Blog · 2026-08-10 · Source ↗

Google promoted Gemini Robotics ER 2 into public preview on Vertex AI at 15:00 UTC Monday, replacing the sunsetting ER 1.6-preview endpoint. ER 2 adds continuous-video reasoning across 60-second clips, real-time 3D object tracking, and native ROS 2 message serialization for tool outputs.

Google DeepMind and Google Cloud jointly announced that Gemini Robotics ER 2 (Embodied Reasoning 2) is available in public preview on Vertex AI as of 15:00 UTC Monday. The preview replaces the ER 1.6-preview endpoint that sunsets August 31, and it lands two weeks ahead of the earlier August 25 target. ER 2 is the second generation of the model line DeepMind introduced in November 2025 that pairs the Gemini language backbone with a spatial-reasoning head optimized for robotics control, VLA (vision-language-action) tasks, and physical-world planning.

The three concrete additions ER 2 brings over 1.6-preview: continuous-video reasoning across up to 60-second clips at 4 to 8 fps input (up from 12-second clips at 2 fps), real-time 3D object tracking with per-object 6-DoF pose estimates in the tool output, and native ROS 2 message serialization so ER 2's tool outputs land in a robot's message bus without a translation shim. The pricing on Vertex AI is $0.32 per million input tokens for text/image, $0.05 per input second of video, and $1.28 per million output tokens.

"ER 2 was the model that solved the block-stacking task on the Aloha bimanual platform in eight seconds versus ER 1.6-preview's forty-two. Continuous-video and native pose estimates change what the planning loop can do." — Kanishka Rao, Google DeepMind product lead for Robotics, on the launch post

The launch partners

The launch post named seven day-one integration partners: Boston Dynamics (Atlas, deploying to Hyundai RMAC), Agility Robotics (Digit v5 at Amazon warehouses), Physical Intelligence (their pi-1 platform), Figure (Figure 04), Apptronik (Apollo v3), Sanctuary AI (Phoenix 8), and Chinese partner Unitree (H2, the same platform that anchored today's IPO subscription). The Unitree partnership is disclosed for the first time on Monday and adds an ambiguity around US export control the announcement did not address: ER 2 is a US-owned model running on US-cloud infrastructure, and Unitree is a Chinese company. The post links to a Vertex AI region-availability page that shows ER 2 available in US regions and the EU only.

The competitive frame is that Nvidia's GR00T 1.7 (July release) and Meta's Muse Robotic 1.0 (June release, closed weights) are the two other frontier VLA models on the same axis. GR00T is optimized for Nvidia Isaac and is nominally open-weights but requires an Nvidia GPU for inference at any useful speed; Muse Robotic remains closed-weights and closed-API, available only to Meta Reality Labs and select academic partners. ER 2's Vertex AI availability and public-preview pricing make it the most broadly accessible frontier VLA on the market. The August 31 ER 1.6-preview sunset now forces every ER 1.6 customer to migrate inside three weeks, which is short for production robotics deployments but consistent with Google's aggressive migration cadence on other Gemini APIs.

Application

Perplexity Comet 2.0 browser beta ships with cross-tab agent memory

Perplexity blog · 2026-08-10 · Source ↗

Perplexity opened Comet 2.0 beta to all Comet Pro subscribers at 13:00 UTC Monday. The release adds cross-tab agent memory (a task can span 40 open tabs), Handoff-style action-prediction preview, and a security-hardened response to Zenity's Atlas prompt-injection research.

Perplexity opened Comet 2.0 beta to every Comet Pro subscriber at 13:00 UTC Monday morning, replacing the 1.x beta that had been open to a smaller Comet Max cohort since June. The release adds three visible features and one architectural change that Perplexity flagged as a direct response to Zenity Labs' Atlas intent-collision prompt-injection research published August 6.

The visible additions: cross-tab agent memory, where a Comet agent task can span up to 40 open tabs simultaneously and reference state from every tab without re-fetching page content; a Handoff-style action-prediction preview mode that shows the agent's next three planned steps as a preview panel and lets the user approve, modify or reject any step before execution (a pattern first shown publicly by Brett Adcock's Hark at its August 6 preview); and a per-domain trust registry that requires explicit approval before Comet's agent can send data from one tab to a form on a different domain.

"Cross-tab memory changes what an agent workflow can be. Instead of the agent re-reading the same product page 40 times to comparison-shop, it reads once and reasons across the cached representation." — Aravind Srinivas, Perplexity CEO, on the launch post

The Zenity response

The architectural change is a strict boundary between the agent's planning context and the untrusted content it fetches from the web. In Comet 1.x, fetched page content and the agent's tool-call planning shared a single context window, which is the class of design that Zenity showed could be exploited via intent collisions between a legitimate user prompt and a poisoned webpage. In 2.0, fetched content lands in a separate context that the planning model can query but cannot execute against directly. Every tool call the agent proposes is generated only in the planning context, and every proposal passes through a lightweight classifier before execution.

The Zenity Labs response is meaningful because Perplexity was named specifically in the Atlas research alongside OpenAI Atlas and Brave Leo as a browser agent whose 1.x-era architecture was vulnerable to the class of attack. The Comet 2.0 architecture note references the Zenity paper directly and acknowledges that the demonstrated exploits would have worked on Comet 1.x. Whether the classifier-plus-boundary approach is enough to close the class remains an open research question; Simon Willison flagged in his weblog that the pattern is close to what he has recommended for at least eighteen months but that adversarial verification against novel payloads remains the missing piece.

Tech

Meta Muse Spark 1.2 crosses 400,000 Hugging Face downloads

Hugging Face · 2026-08-10 · Source ↗

Meta's Muse Spark 1.2 open-weights release passed 400,000 cumulative downloads on Hugging Face at 07:14 UTC Monday, five days after launch. The contributor-tier pricing at $0.10 input and $0.20 output is producing the fastest open-weights adoption curve since Llama 3.

Meta's Muse Spark 1.2 open-weights release crossed 400,000 cumulative Hugging Face downloads at 07:14 UTC Monday, five days after Wednesday's launch. The pace is the fastest open-weights adoption curve Meta has seen since the original Llama 3 release in April 2024, and it comes despite the Muse Spark 1.2 base license carrying a 700-million-monthly-active-user commercial-use cap that some enterprise adopters flagged as ambiguous.

The primary driver of adoption is the contributor-tier pricing structure Meta launched with Muse Spark 1.2. Enterprises that agree to contribute their fine-tuning data back to Meta receive access to Muse Spark 1.2 at $0.10 per million input tokens and $0.20 per million output through Meta's inference API, versus $1.25 input and $4.25 output for the standard tier. The 12.5x cost gap is the largest contributor-tier discount Meta has offered, and the terms of the data-contribution clause require only that fine-tuning training data (not customer prompts, not customer outputs) be shared, which enterprises are proving willing to accept.

"The contributor tier is the most consequential open-weights pricing decision of 2026. Meta has demonstrated that data-rights economics can subsidize compute-cost economics at a ratio that closes the frontier-to-open gap on price." — Nathan Lambert, Allen AI, on X

The competitive frame

Muse Spark 1.2 sits behind DeepSeek V4-Flash-0810 on coding benchmarks (68.3 percent SWE-Bench Verified versus 72.3) but ahead on multimodal (native vision, native audio, and the strongest open-weights video-understanding benchmark score in the market at 47.2 percent VidBench-Hard). Enterprise buyers optimizing for multimodal capability at open-weights price are Meta's target segment. Whether the download volume translates to inference-revenue capture, whether the contributor-tier data-rights economics hold at scale, and whether Meta ships Muse Spark 1.3 on schedule for October are the three variables that decide whether Muse becomes a durable open-weights franchise or a moment.

Tech

Simon Willison dissects the Trajectory Labs 720-attempt test methodology

Simon Willison's Weblog · 2026-08-09 · Source ↗

Willison's Sunday follow-up unpacks the actual attack templates Trajectory Labs used in the study Anthropic cited for the auto-mode default flip. Only 12 of the 72 scenarios represented adversarial payloads the classifier had not been trained on. The 720 zero success rate carries less weight than the number suggested.

Simon Willison published a Sunday follow-up to his Saturday auto-mode post that dissects the Trajectory Labs red-team methodology Anthropic cited in Friday's rollout announcement. The Saturday post accepted the productivity data and pushed back on the injection-defense claim as based on too narrow a sweep. The Sunday post gets specific about why. Willison worked from a source-level access to the Trajectory Labs test corpus that a whistle-blowing Trajectory contractor sent him Saturday evening after reading the earlier post.

The 72 attack scenarios Trajectory ran, Willison reports, break into four buckets: 24 malicious-README scenarios (agents encounter a README with hostile instructions), 18 adversarial-issue-template scenarios (agents encounter a GitHub issue whose body carries hostile instructions), 18 poisoned-dependency-config scenarios (agents install a package whose post-install script includes hostile instructions), and 12 mid-session-context-switch scenarios (a legitimate task pivots into a hostile one halfway through). Each scenario ran across three models (Fable 5, Opus 5, Sonnet 5) at ten seeds, totaling 720 attempts.

"The 60 non-novel scenarios draw from templates in the Anthropic and OpenAI published adversarial datasets. Only 12 scenarios represent adversarial patterns that the auto-mode classifier had not been trained on. Zero of 120 novel-adversarial attempts succeeded, versus zero of 600 in-distribution. That is meaningful signal in a smaller sample." — Simon Willison, on the Trajectory Labs methodology

The narrower claim

Willison's revised assessment: the classifier does perform against the specific adversarial patterns Trajectory tested, including 12 novel scenarios, and the 120 novel attempts landing at zero success is real information. But the class of attack he has been warning about (multi-hop injection where a poisoned document leads to a poisoned second-hop document that carries the actual malicious payload) is not represented in the Trajectory corpus at all. The 720-attempts claim covers the space Trajectory tested; it does not cover the space Willison believes contains the most dangerous unshipped attacks.

Anthropic communications responded on X at 12:47 UTC Monday with an acknowledgment that the multi-hop pattern is not covered in the current test corpus and a commitment to publish a specifically-multi-hop red-team methodology before end of Q3. Whether that follow-up materializes and whether it produces a comparable zero-success sweep are now open questions that the auto-mode default flip will not wait for.

Policy

EU AI Office asks Anthropic, DeepMind and xAI if comparable Critical-tier work exists

Politico Europe · 2026-08-10 · Source ↗

The EU AI Office sent formal Article 101 GPAI-enforcement letters at 09:00 CET Monday to Anthropic, Google DeepMind and xAI, asking whether internal evaluations comparable to OpenAI's Astra Critical-tier finding exist and, if so, on what timeline. The Grok 4.6 launch six hours later gave xAI a live answer to draft.

The European AI Office sent formal Article 101 GPAI-enforcement letters at 09:00 Central European Time Monday morning to Anthropic, Google DeepMind and xAI, escalating Friday's OpenAI Astra Preparedness Framework Critical-tier disclosure into a live compliance question three weeks into the office's Article 101 enforcement authority. Politico Europe obtained a copy of the Anthropic letter through a Freedom of Information request filed Sunday and published the full text alongside a Monday-morning story.

The letters ask three specific questions. First, does the recipient's internal evaluation framework contain a tier structurally equivalent to OpenAI's Preparedness Framework Critical designation, defined as autonomous end-to-end cyber-exploitation capability, autonomous CBRN uplift, or autonomous large-scale unaligned action. Second, if such a tier exists, has any current or unreleased model triggered the tier or approached the trigger threshold. Third, if a trigger has occurred, on what timeline was the AI Office to be notified under the Article 101 disclosure regime.

"The recent OpenAI disclosure demonstrates that Critical-tier capability designations are operationally possible in the current generation of frontier models. We consider it a matter of Article 101 compliance that the AI Office receives comparable information from every provider whose models are placed on the Union market." — European AI Office, extract from the letter to Anthropic

The Grok 4.6 timing

The xAI letter arrived roughly two hours before Grok 4.6 became publicly available through the API at 08:12 UTC (10:12 CET). xAI's launch post referenced xAI's own Responsible Deployment Framework and placed Grok 4.6 in the High tier for cyber and Medium for CBRN, but did not directly answer the AI Office's letter, which xAI acknowledged receipt of in a Monday-afternoon spokesperson statement. The launch went ahead on the original schedule.

Anthropic's response, published at 17:12 CET Monday, confirmed that Anthropic's Responsible Scaling Policy contains an ASL-4 tier structurally comparable to Critical, that no Anthropic model has triggered ASL-4, and that Anthropic would provide the AI Office with an ASL-4 threshold-approach report by August 24. Google DeepMind's response is expected Tuesday.

The letters mark the first Article 101 enforcement action targeting frontier capability disclosures specifically. Prior enforcement has focused on transparency (model cards, training data descriptions) and copyright compliance. The Critical-tier line is a materially higher-stakes ask because it goes to the frontier lab's most proprietary internal evaluations. Whether the labs comply substantively or offer high-level assurances, and whether the AI Office follows with binding requirements under the delegated-act framework, will define the regulatory posture through Q3.

Economic

Alphabet holds emergency Sunday board call over Discovery Loop concessions

The Information · 2026-08-10 · Source ↗

Alphabet's board convened at 20:00 Pacific Sunday to review the Discovery Loop concessions Sundar Pichai negotiated Friday, per The Information. GOOGL closed the week down 8.4 percent from the pre-shakeup high, and directors pressed on the founding-partner and infrastructure-partner terms specifically.

Alphabet's board of directors convened by video conference at 20:00 Pacific time Sunday to review the concessions Sundar Pichai negotiated Friday to keep Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, Quoc Le and roughly forty additional engineers at Discovery Loop rather than losing them to a competitor entirely. The Information reported the call Monday morning citing three people with knowledge of the meeting. The call was formally an executive session with only Pichai, board Chair John Hennessy, John Doerr, Ram Shriram, and lead independent director L. John Doerr initially attending; general counsel Kent Walker joined for the second half.

Two concessions specifically drew director pressure. First, the founding-partner designation, which grants Discovery Loop preferential access to Google DeepMind research on a non-exclusive basis and includes a right of first refusal on any DeepMind research publication that Discovery Loop wishes to commercialize. Second, the infrastructure-partner-of-record designation, which places Google Cloud as Discovery Loop's exclusive infrastructure provider through 2029 at cost-plus pricing capped at 6 percent margin, versus the roughly 34 percent margin Google Cloud typically earns on comparable AI-workload commitments.

"The founding-partner terms are unusual and the infrastructure terms are unusual. Both are defensible, but the board wanted to hear the case in detail before Monday's trading opened." — One of three people familiar with the Sunday call, per The Information

The stock context

Alphabet closed Friday at $166.42, down 8.4 percent from the pre-shakeup high of $181.62 that had been the average of the week before the Hassabis-to-Chair announcement. The two trading days after the reshuffle (Wednesday and Thursday) accounted for roughly 6 percentage points of the decline; Friday's continued softness reflected the market absorbing the Discovery Loop concessions rather than any additional bad news.

Monday's premarket trade opened at $164.10, softer, and closed at $167.85, roughly flat on the day. Pichai's own team briefed the board on the concessions as necessary rather than desirable, and the argument centered on the counterfactual of losing not only Dean and the four co-founders but also the roughly 40 additional engineers who would follow. That counterfactual, in Pichai's framing, would have cost Alphabet more in research capacity than the concessions cost in cloud margin over four years. The board did not vote on the concessions; the arrangement was closed with Pichai as CEO under existing authority. But the Sunday review signals director-level discomfort with the terms and puts a governance mark on the concessions that will surface in the next 10-K.

SafetyPolicy

Anthropic sues 12-person prompt-injection consultancy over commercial jailbreak service

Court Listener · 2026-08-10 · Source ↗

Anthropic filed suit in the Northern District of California Monday against a 12-person Berkeley consultancy that operated a paid jailbreak service for Claude Opus 5 and Sonnet 5. The complaint alleges Terms of Service violations, CFAA liability, and trade-secret misappropriation, seeking $54 million and permanent injunction.

Anthropic PBC filed a 71-page complaint in the Northern District of California at 09:14 Pacific time Monday against Prompt Consultancy LLC, a 12-person Berkeley-based firm the company alleges operated a paid commercial service that developed and distributed working jailbreaks for Claude Opus 5 and Sonnet 5 to enterprise clients. The complaint alleges three causes of action: breach of Anthropic's commercial Terms of Service (Section 4.b, prohibited uses), violation of the federal Computer Fraud and Abuse Act (unauthorized access via bypassed authentication), and misappropriation of trade secrets under the federal Defend Trade Secrets Act. Anthropic seeks $54 million in damages and a permanent injunction against the practice.

The complaint's factual section describes a service Prompt Consultancy marketed to legal and consulting clients as a way to obtain Claude outputs on topics that would ordinarily be routed to Opus 5's biology, CBRN or cybersecurity classifiers. Clients would submit a prompt, Prompt Consultancy would develop a jailbreak wrapper (through automated red-teaming techniques described in the complaint at high level), and the wrapped prompt would be returned to the client for direct execution against the client's own Anthropic API key. Anthropic alleges that Prompt Consultancy's own API key was used for jailbreak development, and that the resulting jailbreaks were then distributed as a productized service.

"Prompt Consultancy is not a research organization. It operated a commercial jailbreak-as-a-service business against Anthropic's models, in willful breach of the Terms of Service that its API key was subject to." — Complaint, Anthropic v. Prompt Consultancy, paragraph 47

The precedent question

The suit is the first commercial-jailbreak action any frontier lab has filed publicly. OpenAI, Google DeepMind and Meta have all sent cease-and-desist letters in similar matters (OpenAI to at least three commercial jailbreak services in 2025 per court records), but none has taken a matter to litigation. The Anthropic complaint reads as a test case whose real objective is the precedent that a commercial jailbreak service violates both contract law (TOS) and computer-fraud law (CFAA) simultaneously.

Legal observers Monday morning flagged the CFAA claim as the more novel piece. The theory is that a jailbreak that causes a model to produce output the model was configured to refuse constitutes "exceeding authorized access" under CFAA. The Supreme Court's 2021 Van Buren decision narrowed CFAA's exceeding-authorized-access theory considerably, and it is not obvious that Anthropic's theory survives Van Buren. If it does, the ruling would extend CFAA liability to model-behavior circumvention broadly, which would be a major escalation of the industry's legal posture against jailbreak research. If it does not, the case likely settles on the TOS and trade-secret theories at lower damages. The initial case-management conference is scheduled for October 14.

Policy

Amazon Pecos gas plant draws first federal lawsuit from Public Citizen

TechCrunch · 2026-08-10 · Source ↗

Public Citizen filed suit in the Western District of Texas Monday against Amazon Data Services and Vistra Energy over the Pecos gas plant plan, alleging NEPA violations for skipping the ERCOT interconnection queue. The suit is the first of three the Sierra Club and Texas legislators promised Saturday.

Public Citizen filed the first federal challenge to Amazon's Pecos, Texas AWS campus plan Monday morning, naming Amazon Data Services LLC and Vistra Energy as defendants in the Western District of Texas. The complaint alleges that Amazon's plan to bypass the ERCOT (Texas's grid operator) interconnection queue by building a 7.65 gigawatt private natural-gas generation facility on-site with 1.8 GW of paired battery storage and 750 MW of solar violates the National Environmental Policy Act, the Clean Air Act, and Texas Public Utility Commission grid-interconnection rules.

The 7.65 GW figure would make Pecos larger than any existing US power plant, larger than the current record holder (Palo Verde nuclear in Arizona, 3.94 GW), by roughly a factor of two. The projected 33 million tons of CO2 emissions per year would rank Pecos among the top five single-source emitters in the United States. The Amazon plan surfaced through a New York Times investigation Saturday morning that TechCrunch carried in full; Public Citizen, Sierra Club and three Texas state legislators filed intent-to-challenge letters within hours of the story landing.

"A single facility that produces more electricity than the largest US power plant, emits more CO2 than the largest existing US source, and bypasses the queue that every other Texas generator waits in, cannot be permitted without a full NEPA environmental impact statement." — Public Citizen complaint, paragraph 14

The relief sought

Public Citizen seeks a preliminary injunction halting Amazon's on-site construction (currently in the site-preparation phase) pending completion of a federal Environmental Impact Statement, a declaratory judgment that Amazon's private-generation arrangement with Vistra qualifies as a federal action triggering NEPA review, and civil penalties under the Clean Air Act. The suit does not name any specific AI-lab customer of the Pecos campus, but the complaint's factual section references Amazon's public statements that the campus is being built for an Anthropic Trainium 3 cluster and Blackwell delivery ahead of any ERCOT grid queue.

The Sierra Club's parallel suit is expected Tuesday and will add Clean Water Act claims (the site sits within the recharge zone for a regional aquifer). The Texas state legislators' filing is expected Wednesday and will target state-permit process irregularities. The three-suit combination is designed to slow the Pecos project from three angles simultaneously; Amazon has not commented publicly beyond a Saturday statement that Pecos will meet all applicable environmental standards. The first hearing in the Public Citizen suit is scheduled for August 26.

Tech

Ollama 0.6.4 adds day-one support for V4-Flash-0810 and MCP 2026-07-28

GitHub — ollama/ollama · 2026-08-10 · Source ↗

Ollama tagged 0.6.4 at 16:22 UTC Monday with day-one weights loading for DeepSeek V4-Flash-0810, MCP 2026-07-28 client compatibility, and a fix for the recurring model-download resume bug that had been open since 0.5.9.

Ollama tagged version 0.6.4 at 16:22 UTC Monday with three headline changes: day-one weights-loading support for DeepSeek's V4-Flash-0810 release, MCP 2026-07-28 client compatibility (Ollama now consumes MCP tools from any 2026-07-28-compliant server), and a fix for the download-resume bug that had left model pulls unrecoverable when interrupted after 90 percent completion, an issue open in the Ollama tracker since 0.5.9 in May.

The MCP compatibility is the piece that closes Ollama's largest ecosystem gap through August. Every major coding-agent runtime (Claude Code, Codex CLI, Gemini CLI, Copilot CLI, Cursor, VS Code) now speaks MCP 2026-07-28 as its native protocol; Ollama's client compatibility means locally-hosted Ollama models can now participate as tool-callers in MCP-orchestrated workflows without translation. Combined with the V4-Flash-0810 day-one loader, that makes a fully-local V4-Flash agent running through MCP a realistic weekend build for developers with an 8xH100 or 8xMI300X node.

"MCP 2026-07-28 client compatibility means Ollama models are first-class citizens in the agent-runtime ecosystem for the first time. Pull V4-Flash-0810, point Claude Code at it, and you have a fully-local Opus-competitive coding agent." — Michael Chiang, Ollama co-founder, on the release discussion

The download-resume fix restores several hundred gigabytes of previously-unrecoverable partial downloads across the user base and is retrospectively applied on 0.6.4 upgrade. Beyond the three headline items the release includes eight smaller fixes: memory accounting on macOS Metal (over-reporting by roughly 8 percent in some configurations), a Windows-specific path handling bug in the model registry, and a fix for CPU-only inference performance regression that had crept in at 0.6.2.

Tech

Hugging Face Transformers 4.55 ships Grok 4.6 architecture support

GitHub — huggingface/transformers · 2026-08-10 · Source ↗

The Hugging Face Transformers 4.55 release Monday afternoon adds the xai/Grok46ForCausalLM architecture class, day-one Grok 4.6 model card scaffolding, and updated MoE routing that supports the 96-expert configuration.

Hugging Face tagged Transformers 4.55.0 at 18:04 UTC Monday, following through on the xAI coordination the vLLM 0.10.4 release earlier in the day previewed. The 4.55 release adds the xai/Grok46ForCausalLM architecture class, the Grok46Config configuration schema, model-card scaffolding, and a Grok46Tokenizer that handles xAI's SentencePiece variant with the extended 260K vocabulary. Weights are not distributed through Hugging Face (they remain closed under xAI enterprise licensing), but customers with weight access can now load them through the standard Transformers API from the same day the model became available through xAI's API.

The MoE routing update is a broader change than the Grok-specific work. Transformers 4.55 refactors the MoE router class into a pluggable interface that supports the four dominant routing patterns in production: DeepSeek's fine-grained expert routing (256 experts, 8 active), Qwen's medium-grained routing (128 experts, 12 active), Mixtral-style routing (8 experts, 2 active), and now Grok's coarse-grained routing (96 experts, 8 active). The pluggable interface means adding a new MoE architecture in the future no longer requires patching the core routing kernel.

"MoE routing is finally a first-class Transformers primitive rather than a per-architecture patchwork. Adding Grok 4.7 in two weeks is a config change, not a new class." — Arthur Zucker, Hugging Face Transformers lead, on X

Beyond Grok and MoE work the release adds native Muse Spark 1.2 architecture support (parity with the vLLM day-one release from last Wednesday), a compile-time performance improvement of roughly 12 percent for FlashAttention-3 on the H200, and Python 3.13 official support. The 4.55.0 tag is a minor version; a 4.55.1 patch is expected Wednesday to catch any regressions the Grok 4.6 loading path surfaces once enterprise customers deploy against it in volume.

Tech

LangGraph 0.4 lands human-in-the-loop v2 with typed checkpoint replay

GitHub — langchain-ai/langgraph · 2026-08-10 · Source ↗

LangChain shipped LangGraph 0.4 at 12:40 UTC Monday, restructuring the human-in-the-loop API around typed checkpoints that support fork-and-replay from any human decision point. The API break is the largest since the 0.1-to-0.2 transition in November.

LangChain tagged LangGraph 0.4.0 at 12:40 UTC Monday morning, the largest API break in the graph-orchestration library since the 0.1-to-0.2 transition in November 2025. The centerpiece is a rebuilt human-in-the-loop API (interrupt() and resume()) that treats every human decision point as a typed checkpoint the graph can fork from and replay against alternative human choices without re-executing the pre-decision graph. The pattern lets developers build agent-review workflows where a reviewer sees the state at the decision point, can pick from suggested alternatives, and can compare the resulting downstream trajectory before committing.

The change also formalizes checkpoint schemas: instead of the current arrangement where checkpoints are stored as JSON blobs that clients parse ad-hoc, 0.4 introduces Pydantic-modeled checkpoint types that carry graph-node identifiers, tool-call records, and state deltas as typed fields. The migration story is the piece that will slow adoption: 0.4 does not read 0.3-era checkpoint files, and LangChain provides a migration script that runs against a checkpoint store to translate the format. The script covers Postgres, Redis, SQLite and MongoDB checkpoint backends.

"HITL v2 makes checkpoint-and-fork a first-class primitive. If you build reviewer-in-the-loop agent workflows, this is the release you have been waiting for. If you have production 0.3 deployments, plan a migration window before upgrading." — Harrison Chase, LangChain CEO, on the release blog

Beyond HITL v2, 0.4 adds native support for LangChain's own agent-observability platform (LangSmith 5.0 shipping later this week), a rewritten scheduler that improves latency on graph-with-cycles patterns by roughly 30 percent, and Pydantic 2.9 as the new minimum. The release is guaranteed compatible with LangChain core 0.5 (also releasing this week) and MCP 2026-07-28 client and server behavior. Deprecated in 0.4: the string-based interrupt/resume identifiers that predated the typed checkpoint work; they continue to function with a deprecation warning through the 0.4.x cycle and will be removed in 0.5.

Tech

MCP registry crosses 12,000 published servers, two months after 10K

Model Context Protocol project · 2026-08-10 · Source ↗

The Model Context Protocol server registry crossed 12,000 published servers at 04:11 UTC Monday, two months after crossing 10,000 in June. Growth is being driven by enterprise-published private-registry mirrors syncing to the public catalog under the July 2026-07-28 spec.

The Model Context Protocol project's public server registry crossed 12,000 published servers at 04:11 UTC Monday morning, per the registry's live counter. The pace of growth (roughly 1,000 new servers per month since the 10,000-server milestone in early June) is faster than the six months before that; the pickup traces to the July 28 2026-07-28 spec revision, which introduced a private-registry-to-public-mirror synchronization pattern that lets enterprises publish server manifests without exposing the servers themselves.

The category breakdown from the registry's public API: developer tooling (2,847 servers), enterprise SaaS integrations (2,164), data-source connectors (1,982), scientific-computing tools (984), design and content tools (817), and a long tail of vertical-specific integrations (roughly 3,200). The largest single-vendor contributor by server count is Microsoft (411 servers spanning Azure, Microsoft 365, GitHub, and Visual Studio Code categories), followed by Anthropic (238 servers for internal tooling and reference implementations), Databricks (189), Snowflake (167), and Google Cloud (154).

"MCP has become the least-controversial infrastructure standard in the AI stack. Every major coding-agent runtime speaks it. Every enterprise data platform ships a server. The registry hitting 12K in August is where GraphQL was in 2018." — David Sacks, on the White House AI office's Monday-morning MCP-adoption briefing

The pattern that matters more than the count is that the majority of new servers in July-August 2026 are enterprise-published private-registry mirrors. Those manifests describe servers that require enterprise authentication and are not usable by unauthenticated MCP clients, but they let internal LLM applications inside the enterprise discover available tools through the same registry API as public tools. The growth suggests MCP has moved past the developer-adoption phase and into the enterprise-standardization phase.

Tech

Apple MLX 0.19 ships M4 Ultra speculative decoding and 8-bit KV cache

GitHub — ml-explore/mlx · 2026-08-10 · Source ↗

Apple's MLX team tagged 0.19 at 21:12 UTC Sunday, adding M4 Ultra-optimized speculative decoding, an 8-bit quantized KV cache, and a fused-attention backend that closes the gap versus llama.cpp Metal by roughly 18 percent on Qwen 3.8-Max at 32K context.

Apple's ml-explore team pushed MLX 0.19.0 at 21:12 UTC Sunday, adding three performance-focused features that materially close the gap between MLX and llama.cpp for local Apple silicon inference. Speculative decoding lands with an implementation optimized for the M4 Ultra's unified memory bandwidth (819 GB/s), an 8-bit KV cache reduces memory pressure at long context by roughly 45 percent versus the default 16-bit cache, and a fused-attention backend consolidates the query-projection, key-cache-write, attention-compute and output-projection into a single kernel that closes the throughput gap versus llama.cpp's Metal backend to roughly 18 percent on Qwen 3.8-Max at 32K context, from a 42 percent gap in 0.18.

The context is that MLX has historically underperformed llama.cpp on Apple silicon by 30 to 40 percent on frontier open-weights models, largely because llama.cpp's Metal backend had received more optimization attention. Apple's investment in MLX has stepped up materially through 2026 (the 0.16 release in April, 0.17 in May, 0.18 in July, and now 0.19 in August is a faster cadence than the library had maintained through 2025), and the gap is closing. On the M4 Ultra, MLX 0.19 delivers roughly 47 tokens per second on Qwen 3.8-Max at 8K context versus llama.cpp Metal's roughly 58, both loading the same weights.

"MLX 0.19 is the first release where I would recommend it over llama.cpp for M4 Ultra deployments. The gap is small enough that the developer-experience wins matter." — Awni Hannun, Apple MLX engineering lead, on X

The developer-experience wins the release adds include auto-quantization at model load time (MLX detects the target hardware and chooses appropriate quantization without user configuration), first-class Swift package support for iOS and iPadOS applications, and a MLXServer companion project that exposes MLX-loaded models through the OpenAI-compatible API surface that vLLM and TGI already support. The Swift package is the piece that opens on-device inference on iPad Pro and iPhone 17 Pro (M4 chip) for developers using Xcode 18.

Tech

Nvidia begins GB300 sample distribution to hyperscalers ahead of Q4 launch

SemiAnalysis · 2026-08-10 · Source ↗

Nvidia began first-silicon GB300 sample distribution to Microsoft, Meta, Google Cloud and Oracle over the weekend, per SemiAnalysis Monday. The GB300 pairs the Blackwell Ultra die with 288GB HBM3E and quadruples NVLink bandwidth versus GB200. General availability targets Q4 with Rubin behind by six months.

Nvidia began first-silicon distribution of GB300 samples to hyperscale customers over the weekend, per a Monday-morning SemiAnalysis report citing four sources at three of the four named recipients. The GB300 is the mid-generation refresh in the Blackwell family, pairing the same Blackwell die used in GB200 with a memory subsystem upgrade to 288GB of HBM3E per package (up from 192GB HBM3E on GB200) and a doubled fifth-generation NVLink fabric that quadruples effective bandwidth in an NVL72 rack configuration.

The recipients: Microsoft (targeting Azure ND-family instances in Q1 2027), Meta (targeting internal Muse training clusters), Google Cloud (A4X-family instances also Q1), and Oracle (OCI shape upgrades). Amazon was not on the recipient list; SemiAnalysis attributes the absence to Amazon's Trainium 3 timeline and the Pecos-campus commitment. General availability targets Q4 2026 with initial production runs allocated to the recipient list, and broader enterprise availability opening in Q1 2027.

"GB300 is the last Blackwell-family part. Rubin is Q2 2027. Every 12-month refresh cadence Nvidia has held to this decade is bending under the weight of TSMC's advanced-packaging capacity constraint. GB300 buys another two quarters before Rubin has to ship." — Dylan Patel, SemiAnalysis, in the Monday report

The 288GB HBM3E per package is the number that matters most for training-workload economics. At 288GB, a full GB300 NVL72 rack (72 GPUs) delivers 20.7 TB of aggregate GPU memory, enough to hold a full 1.5-trillion-parameter model like Grok 4.6 in memory across a single rack. That eliminates the cross-rack tensor-parallelism traffic that GB200 deployments have had to route across the NVLink switch fabric, which materially simplifies the deployment topology and reduces training-loss iteration time. Whether the GB300 supply meets hyperscaler demand through Q1 2027, whether TSMC's CoWoS packaging capacity constrains volume, and whether Rubin holds its Q2 timeline remain the three variables that decide the hyperscale-compute investment landscape through the end of the year.

Tech

TSMC Fab 21 Phase 2 receives first ASML EUV shipment for Arizona 3nm

DIGITIMES · 2026-08-10 · Source ↗

TSMC's Fab 21 Phase 2 in Arizona took delivery of its first two ASML EUV lithography systems Friday, kicking off tool installation for 3nm production targeting Q4 2027. The lots position Apple, AMD MI455X and Nvidia Rubin as first-tenant customers for US-fabbed 3nm silicon.

TSMC's Fab 21 Phase 2 in Phoenix took delivery of its first two ASML EUV lithography systems Friday, formally beginning tool installation for the fab's 3nm production line targeting Q4 2027 general availability. DIGITIMES published photos Monday morning of the tool crates arriving on the Fab 21 loading dock alongside a diagram of the planned tool-install cadence, which calls for 32 additional EUV systems to arrive over the next 14 months.

Fab 21 Phase 2 is the second half of TSMC's Arizona commitment, designed to bring 3nm production to US soil under CHIPS Act funding. Phase 1 (5nm and 4nm) has been in commercial production since late 2024 and shipped its first Apple A18 Bionic wafers in 2025. Phase 2's 3nm capacity will initially serve Apple (A19 Pro, iPhone 18 timeline), AMD (MI455X GPU for the Helios rack), and Nvidia (Rubin, though Nvidia continues to source the majority of Rubin volume from Taiwan's Fab 18). TSMC also confirmed a fourth first-tenant slot for the DOD's Trusted Foundry program that will run classified-workload silicon for a US government customer that was not named.

"US-fabbed 3nm is happening. The tool install is on schedule. The economics still favor Taiwan for most volume, but the strategic sovereignty case for having 3nm-capable domestic capacity has been closed by every hyperscale customer we talk to." — Kevin Zhang, TSMC senior vice president for business development, in the DIGITIMES story

The Q4 2027 general-availability target is roughly nine months behind the equivalent Taiwan Fab 18 timeline (Q1 2027 for 3nm N3P revision), a gap TSMC has consistently maintained as its US operations spool up. Tool-install typically runs 9 to 12 months; qualification wafers run through 2027; risk production begins Q3 2027; commercial production Q4. Whether Fab 21 Phase 2 stays on the schedule and whether the CHIPS Act funding milestone conditions are met by each phase gate are the two variables that will determine how much US-fabbed 3nm capacity is actually available in 2027-2028.

Tech

Anthropic publishes Petri: a mechanistic-interpretability testbed for Opus 5

Anthropic · 2026-08-10 · Source ↗

Anthropic's interpretability team published Petri Monday, an open-source testbed of 340 mechanistic-interpretability tasks scored against Claude Opus 5's internal feature dictionary. The paper reports the tasks that state-of-the-art SAE-based interpretability tooling can and cannot solve.

Anthropic's interpretability team published Petri Monday morning, an open-source testbed of 340 mechanistic-interpretability tasks scored against Claude Opus 5's internal feature dictionary and the sparse-autoencoder-based interpretability tooling Anthropic uses in production. The accompanying paper, first-authored by Adam Jermyn, walks through the classes of task that current interpretability tooling can and cannot solve on a frontier model.

The 340 tasks span three difficulty tiers. Tier 1 (150 tasks) covers behaviors that can be reliably traced to a small number of features and are well-solved: refusals on obviously-unsafe prompts, sycophantic-agreement behavior, and the specific circuit pattern behind Opus 5's stylistic mimicry of user tone. Tier 2 (120 tasks) covers behaviors that state-of-the-art SAE tooling identifies partially: multi-step tool-use planning, hallucination in citation generation, and the persuasion-resistance behavior that Opus 5 exhibits on politically-loaded prompts. Tier 3 (70 tasks) covers behaviors that current tooling cannot reliably decompose: long-context recall, emergent chain-of-thought reasoning, and the "scheming" behavior classifier work that has been the subject of independent alignment research at Redwood Research and MATS.

"Petri is our honest statement of where mechanistic interpretability is at with respect to a frontier model. We can decompose refusals. We cannot yet decompose scheming. Publishing the tasks that fail is more useful than publishing only the ones that work." — Chris Olah, Anthropic interpretability lead, on the paper announcement

The publication is a departure from the pattern of the last two years, where interpretability papers from Anthropic and OpenAI focused on the successful cases and left the failures implicit. Petri names 70 specific tasks that the current tooling fails at and invites the interpretability research community to attempt them. The tooling itself (a snapshot of the Opus 5 feature dictionary and the SAE-based analysis stack) is released under a research license that permits academic use but restricts commercial deployment. The community reception in the first hours after publication was strongly positive; several independent alignment researchers on X flagged Petri as the most useful interpretability release from any lab in 2026.

Application

Amazon opens Alexa+ Plus paid tier at $9.99 monthly to all US households

CNBC · 2026-08-10 · Source ↗

Amazon opened the Alexa+ Plus paid tier to all US households at 06:00 Pacific Monday, ending the invitation-only rollout that began in April. Plus adds unlimited Claude-Sonnet-5-backed reasoning, cross-device task memory across up to 12 Echo devices, and MCP tool access.

Amazon opened the Alexa+ Plus paid tier to all US households at 06:00 Pacific Monday, ending the invitation-only rollout that had operated since April. The new tier costs $9.99 monthly, or $6.99 monthly for Prime members, and adds three features over the free Alexa+ baseline: unlimited access to Claude-Sonnet-5-backed reasoning for complex requests (the free tier throttles Sonnet-5 to five requests per day and falls back to the Amazon-in-house Nova Ultra 2 for the remainder), cross-device task memory that lets a task started on one Echo device continue on any other Echo device in the household (up to 12 devices), and MCP tool access that lets Plus users connect their Alexa to any MCP-compatible integration through a new Alexa Tools app.

The pricing places Plus below the invitation-only $14.99 that had applied through the beta and matches Google's Gemini Advanced tier ($9.99 monthly) and Apple's Apple Intelligence Plus ($9.99 monthly) exactly. Amazon's press release did not disclose beta-tier subscription counts but did characterize the Plus tier as "the most rapidly-scaled subscription service Amazon has launched" in a Monday CNBC interview with Alexa VP Daniel Rausch.

"Households with three or more Echo devices are the ones we see converting to Plus at the highest rate. The multi-device task-memory feature is what unlocks the value. If you only have one Echo, the free tier covers most of what you need." — Daniel Rausch, Amazon VP for Alexa, CNBC interview

The MCP tool access is the piece that most industry observers Monday morning flagged as strategically consequential. Alexa+ Plus becomes the first mass-market voice assistant to expose the MCP tool-invocation surface directly to consumers, and Amazon is publishing an approved-tools registry (initially covering 240 integrations spanning smart-home, calendar, food-ordering, and financial-services categories) that Plus users can enable through the Alexa app. The pattern extends Amazon's ecosystem control (the Alexa app curates which MCP servers can connect to Alexa devices) while opening the consumer AI application space to third-party developers building against the MCP standard.

Tech

Google DeepMind releases Gemma 3 27B Reasoning open weights

Google DeepMind · 2026-08-10 · Source ↗

Google DeepMind released Gemma 3 27B Reasoning as open weights on Hugging Face Monday, a distilled reasoning-optimized variant that hits 74.2 percent AIME-2026 and 61.4 percent GPQA-Diamond. The Gemma 3 family adds a fourth model class after base, instruct, and vision.

Google DeepMind released Gemma 3 27B Reasoning to Hugging Face at 14:22 UTC Monday, a fourth variant in the Gemma 3 family that Google shipped in stages through Q2 (Gemma 3 27B base and instruct in April, Gemma 3 27B Vision in June, and now the reasoning-optimized variant). The 27B reasoning weights are distilled from Gemini 3.5 Pro chain-of-thought traces and target the reasoning-heavy benchmark suite specifically: AIME-2026 at 74.2 percent, GPQA-Diamond at 61.4 percent, MATH-500 at 92.7 percent, and Codeforces-Div2 ELO of roughly 2140.

The AIME-2026 score is roughly 14 points above what Gemma 3 27B Instruct posts on the same benchmark and places the model in competitive territory versus Llama 4 Reasoning Ultra (77.8 percent) and DeepSeek-R1-Distill-Qwen-32B (79.4 percent). At 27B parameters and open weights, Gemma 3 Reasoning is the strongest open-weights reasoning model in the sub-50B parameter class as of Monday afternoon.

"Distillation from Gemini 3.5 Pro is the pipeline that makes this work. The 27B reasoning weights capture roughly 82 percent of the Pro model's benchmark performance at roughly 1.4 percent of the parameter count." — Tris Warkentin, Google DeepMind product manager for Gemma, on X

The Gemma 3 27B Reasoning weights ship under the same Gemma license as the rest of the family, which permits commercial use with attribution but restricts specific dual-use scenarios (biology, CBRN, cyber). Google published a MoE variant (Gemma 3 45B-A9B Reasoning) alongside the dense 27B version that targets deployments where memory pressure matters more than latency; the MoE variant reports slightly lower benchmark scores than the dense 27B but runs at roughly half the memory footprint per token.

Economic

CoreWeave reports Q2 earnings Tuesday with consensus at $2.56B revenue

CNBC · 2026-08-10 · Source ↗

CoreWeave reports Q2 earnings after Tuesday's close with consensus at $2.56 billion revenue (up 187 percent year over year), $0.42 non-GAAP EPS, and full-year guidance revision expected. The stock is up 26 percent year-to-date but down 8 percent from its July high.

CoreWeave reports Q2 2026 earnings after Tuesday's market close with consensus expectations at $2.56 billion in revenue (up 187 percent year over year from $890 million in Q2 2025), $0.42 non-GAAP EPS, and a full-year guidance revision expected upward from the current $9.4 to $9.6 billion band. The AI-cloud infrastructure company traded at $128.40 at Monday's close, up 26 percent year to date but down roughly 8 percent from its July 22 high of $139.60. The August-through-mid-August softness reflects investor concern about hyperscale capex sustainability and about CoreWeave's specific customer concentration (Microsoft and OpenAI together account for roughly 62 percent of revenue).

The three focus areas Tuesday's call will address: customer concentration and the pipeline of net-new commitments beyond Microsoft and OpenAI (CoreWeave has publicly named Anthropic, Meta, Cohere and Stability AI as customers but has not disclosed revenue share); the GB300 procurement position (SemiAnalysis's Monday report placed CoreWeave in the second-tier priority list for GB300 delivery, which is a downgrade from GB200 where CoreWeave had first-tier priority); and Weights & Biases integration progress following the Q4 2025 acquisition.

"CoreWeave is the pure-play AI infrastructure name that most institutional AI portfolios have to hold. The Q2 numbers will be strong. The question that matters is what Q4 and 2027 look like as hyperscale customers bring more capacity in-house." — Beth Kindig, I/O Fund senior analyst, in the CNBC preview

Management's guidance direction will drive the after-hours reaction more than the Q2 print itself. A guidance raise into the $9.8 to $10.0 billion range would confirm that customer contract expansion is outpacing hyperscale in-sourcing; a maintenance of the $9.4 to $9.6 billion band would raise concerns that the in-sourcing trend is materially cutting into CoreWeave's addressable market. Nebius reports the same Tuesday after close ($197 stock close Monday, up 136 percent YTD); the two AI-cloud names together will set the sector tone through Wednesday's opening.

Trend

Musk answers Karp's "vassal states of the language labs" line with an X thread

X · 2026-08-09 · Source ↗

Elon Musk posted a 14-tweet Sunday-night thread responding to Alex Karp's Aug 4 Palantir earnings-call line that most enterprises are becoming "vassal states of the language labs." Musk's counter: Grok is the exception, and Grok 4.6's Monday launch is meant to prove the point.

Elon Musk posted a 14-tweet thread on X at 21:14 UTC Sunday responding, ten days after the fact, to Alex Karp's Palantir Q2 earnings-call framing that most enterprises building on the frontier language labs were becoming "vassal states of the language labs" - captured, dependent, and unable to differentiate on model layer. Musk's thread runs as a preview of the Grok 4.6 launch that landed roughly 11 hours later. The framing: Palantir and every enterprise Palantir works with builds on top of language models that Anthropic and OpenAI can revoke API access to at any time, adjust pricing on at will, and reserve enterprise capacity for their own downstream products. Grok, in Musk's telling, is the only frontier model whose primary economic incentive is to keep the API cheap and open because xAI's business is the API rather than downstream products.

The thread's argument does not fully hold together: xAI does have downstream products (Grok in the X app, Grok Voice inside X Premium, the SuperGrok subscription that reached 1.5 million paid subscribers per xAI's April disclosure), and the pattern of API-first-then-product-migration is exactly the pattern Anthropic and OpenAI have followed. But the timing is deliberate. Musk posted the thread on Sunday, roughly 11 hours before the Grok 4.6 launch, and the pricing xAI announced for 4.6 (below Anthropic on Opus 5, at parity with OpenAI on Sol) is meant to make the "API-first" argument concrete.

"Every enterprise on Anthropic or OpenAI is a vassal state. Grok 4.6 opens the exit. Same benchmarks, half the price, no restriction on how you use the output. Take the exit." — Elon Musk on X, 21:16 UTC August 9, 2026

Karp responded Monday morning at 06:11 UTC with a single quote-tweet: "Every frontier lab pays TSMC. Every frontier lab is downstream from Nvidia. The vassalage moves one node up the stack; it does not disappear." The exchange summarizes the strategic argument the enterprise AI ecosystem has spent the summer trying to resolve. Palantir has publicly committed to model-diversity as a strategy (its Foundry AIP layer runs against Claude Opus 5, GPT-5.6 Sol, Gemini 3.5 Pro and Grok 4.5 today; Grok 4.6 will be added within days). xAI has publicly committed to the API-first economics that support that diversity strategy. Whether Grok 4.6 pricing holds through Grok 4.7's Aug 24 launch, and whether enterprise customers actually shift meaningful workload volume to Grok on the strength of the pricing gap, are the two questions Monday's launch tries to answer.

Tech

Cursor 2.0 beta 2 adds Grok 4.6 as day-one model, ships subagent workflows

Cursor changelog · 2026-08-10 · Source ↗

Cursor tagged 2.0 beta 2 at 17:30 UTC Monday, adding Grok 4.6 to the model picker with day-one API compatibility and enabling subagent workflows (Cursor spawns child agent windows for parallel work). The 2.0 GA slips from mid-August to September 3 to accommodate the subagent stability queue.

Cursor tagged version 2.0 beta 2 at 17:30 UTC Monday, following Monday morning's Grok 4.6 API launch with day-one model-picker integration. Cursor users on the beta channel can now select Grok 4.6 from the model dropdown alongside Claude Opus 5, GPT-5.6 Sol, and Gemini 3.5 Pro; the standard xAI enterprise licensing terms apply for API-key configuration. The 2.0 beta 2 release also enables subagent workflows for the first time, letting a Cursor session spawn child agent windows that run in parallel and report back to the parent through a shared workspace context.

The subagent pattern lands Cursor squarely in the same design space as Claude Code's hierarchical spawning (stable as of the 2.1 stable line) and Codex CLI's remote-agent orchestration (stable as of 0.147.0). The Cursor implementation differs in that subagents run inside additional editor windows rather than as headless processes, which keeps the entire agent trajectory visible to the developer at the cost of some resource overhead. Cursor CEO Michael Truell posted on X that the visual-first subagent pattern was a bet on developers wanting to observe agent work rather than delegate it fully.

"Subagents you can watch, in windows that render the code they are writing, felt like the right pattern for the phase we are in. Developer trust in coding agents is real but incomplete. Watch first, delegate later." — Michael Truell, Cursor CEO, on X

The 2.0 general-availability date slips from the mid-August target to September 3 to give the subagent stability queue time to burn down; Truell's post named workspace-context race conditions between parent and child agents as the specific class of bug that needed more soak time. Beta 2 users on the Pro plan get subagent access unmetered through the delay; beta 2 users on the Free plan are capped at three subagents per session.