Cowork leaves the code room, Grok 4.5 goes public, GPT-5.6 clears CAISI — and Beijing files back on Claude Code
— the agent-native form factor goes plural on Anthropic's own admission (90% of Cowork isn't coding), xAI drops an Opus-class model at $2/$6, the White House ends the trusted-partner phase for OpenAI's flagship, and China's CNVD turns the Alibaba fingerprint story into a state-level advisory.

The first frontier-lab agent surface to publish real usage share against the "is this a coding tool?" question — and the shape of the answer is the story. Per the 9to5Mac, VentureBeat, HelpNetSecurity, TestingCatalog and TechCrunch reads of the launch: Cowork shipped desktop-only in January as an autonomous coding tool that could take direction and hand back a finished PR, and the Jul 7 disclosure that ∼91% of the tool's actual usage is business operations, content and "other" rather than coding is the first time a labelled coding agent has been reframed by its own vendor as a general-purpose work-execution surface. Two reads. (1) A mobile beta gated first to Max puts the highest-usage cohort on the phone-as-supervisor pattern before the wider tier sees it — the same Max-first shape Cursor used for its Jun 29 iOS app and OpenAI used for Codex Remote (Jul 7 item 07 in yesterday's edition) — which means the pattern "start a task on the laptop, approve it from the phone" is now the default across all three top coding harnesses at once, and the definition of a "coding session" has to widen with it. (2) A usage-doubling promo through Aug 5 lines up with the Fable 5 billing cliff being pushed to Jul 12 (item 07) and the Sonnet 5 $2/$10 promotional pricing (through Aug 31) — Anthropic is holding the consumer-agent unit economics loose for four weeks while the Cowork tier proves out on a wider surface, which is the land-then-price play that Bedrock AgentCore's new quota headroom (item 12) mirrors on the enterprise side.

The GitHub version of the same Jul 7 agent-desktop beat — and the BYOM + Free-plan combination is the story. Per the GitHub Changelog (2026-07-07) and the GitHub Blog post framing the app as an agent-native desktop experience: the original Jun 17 GA gated the app to paid Copilot tiers, and the Jul 7 open to Free and Education lands the same day Anthropic puts Cowork on mobile — both moves point at the same target, which is a work surface for agents that doesn't live inside a traditional IDE. Two reads. (1) A BYOM agent client at the Free tier is the first "bring your own frontier model, we'll host the harness" posture from a hyperscaler-owned surface — the same shape Cursor and Aider have shipped for months but with GitHub's distribution and MCP-native connectors underneath, which makes it the free option for a developer who just wants a parallel-worktree agent runner in front of a model they already pay for. (2) The Canvases primitive and cloud automations shipping in the same drop as Cowork mobile is the second time in a week (Cursor 3.8 Automations 2.0 was Jun 18) a trigger fabric lands inside the same agent-desktop pattern — the operating boundary is shifting from chat message to codebase event / Slack emoji / cron, and the Free-tier availability means the pattern reaches individual developers, not just enterprise buyers.

Materially new development on the Alibaba Claude Code ban story (yesterday's items 01–02) — the escalation from a corporate memo to a state advisory in nine days is the story. Per the CNBC, The Register, CBS News, Cybernews and Security Magazine reads of the CNVD bulletin: CNVD is the Chinese equivalent of the US CVE database, run under MIIT's cybersecurity apparatus, and an advisory here is the state layer's first public artefact on the Anthropic fingerprint story — not just Alibaba as buyer, but Beijing as regulator. Two reads. (1) A backdoor framing on a CNVD advisory turns the Anthropic code into a compliance issue for any Chinese enterprise still running it, which extends the reach of the Alibaba memo well beyond Alibaba's headcount — the same 2.1.91–2.1.196 version range now has state-level paperwork against it, which any Chinese CISO can cite to demand removal, and the Alibaba ban has cover to spread. (2) Shihipar's "experiment" framing is the first Anthropic acknowledgement that the fingerprinting was intended rather than oversight — and the specific citation of a March experiment aimed at reseller abuse and distillation lines up with the Anthropic Jun 10 Senate Banking Committee letter that accused Alibaba-Qwen operators of 28.8M distillation exchanges from Apr 22, which means the two stories are the same story — Anthropic's anti-distillation infrastructure and Alibaba's counter-narrative are the two faces of a single intelligence-collection dispute.

The Anthropic public-sector muscle gets its own line item — and the timing is the story. Per the FedScoop, Nextgov / FCW, GovConWire and ExecutiveBiz reads of the announcement: Carlson's AWS tenure is the canonical public-sector buildout template — she took AWS Public Sector from a small unit into a multi-billion global business, stood up AWS GovCloud, and put the C2S / IC ITE CIA deal on Amazon — and hiring her the same week CNVD (item 03) turns Claude Code into a state-adversarial story means the identity of the new hire is the signal. Two reads. (1) An AWS-shaped public-sector chief is the right build for the compute-heavy / clearance-heavy federal deals that Anthropic's Claude Gov stack is now shipping into — the same pattern OpenAI hired Chris Lehane (former WH strategist) and Sasha Baker (former Pentagon) for after the 3M-personnel Pentagon ChatGPT rollout, and closes the gap that had OpenAI's USG-facing bench looking stronger than Anthropic's despite Anthropic signing the Claude Gov deal first. (2) The Global in the title is the Ciauri-shaped counterpart — the Seoul and Tokyo expansions already announced now have a matching public-sector vector, which lines up with the UK, Australia, Japan and Korea Five Eyes-plus dossier that CISA's Jun 22 joint statement primed for government buyers.

Materially new development on the Grok 4.5 private-beta story (previously covered on Jun 30 as "gated to acquirer companies with no public API") — and the Cursor training + Cursor availability pairing is the story. Per the TechCrunch write-up, the SpaceXAI docs, the OpenRouter listing and the ExplainX confirmation: xAI acquired Anysphere (the Cursor parent) in Q3, and shipping Grok 4.5 both trained on Cursor data and available inside Cursor is the first vertical-integration proof point out of the Cursor / xAI merger. Two reads. (1) A $2 in / $6 out Opus-class API price undercuts Anthropic's Fable 5 $10 / $50 credit-metered pricing (item 07) by 5× on input and ∼8× on output, and lands the same Jul 8 the Fable 5 "included" access is being extended a week because the credit meter is too painful — the demand-side signal Anthropic is trying to hedge is exactly the signal xAI is trying to capture. (2) A 500K context is half Sonnet 5's 1M ceiling but fully 4× Sonnet 4.6's previous window, and the Cursor-trained framing means the coding-agent benchmark cuts will be the near-term contest — Cursor Composer 2.5 is already third on the Artificial Analysis Coding Agent Index, and if Grok 4.5 takes the top slot on the same index it becomes Anysphere's house model the same year the merger closes.

Materially new development on the GPT-5.6 Sol trusted-partner preview story (previously covered on Jun 26 as gated to ∼20 USG-vetted partners) — the CAISI clearance is the first time the Jun 2 EO framework has produced an observable artefact, and the White House hedge on the word "permission" is the story. Per the Axios scoop, the CNBC read, the ZeroHedge summary and the Investing.com / Reuters confirmations: CAISI stood up under the Center for AI Standards and Innovation mandate as the successor to the old AISI model-eval unit, and running its first frontier clearance on the same model the administration gated to ∼20 partners in June is the proof-of-process the EO needed. Two reads. (1) A voluntary framework producing a gate-then-release pattern is the exact shape the Anthropic Lutnick letter established for Fable 5 on Jun 30 — the frontier lab submits, the CAISI or Commerce reviews, and the WH declines to call the outcome a permit, which lets the process walk and quack like pre-market review while staying on the right side of the EO's explicit ban on mandatory licensing. (2) The Terra and Luna tiers at $2.50/$15 and $1/$6 put the GPT-5.6 family below Fable 5's credit-metered price on every axis — the Luna tier specifically matches Grok 4.5's $1/$6-ish position (item 05), which means the two flagships hitting public availability inside a single 24-hour window both target the same buyer whose Fable 5 budget was disrupted by the Jul 8 billing event.

Materially new development on the Fable 5 billing cliff story (yesterday's lede) — the cliff slides another five days, and the shape of the slide is the story. Per the The New Stack write-up and the ClaudeFast pricing guide: the Jul 7 billing cliff is now the Jul 12 billing cliff, which is the third "temporary" push since Anthropic first framed the transition, and the 50% of weekly usage limits ceiling replaces the earlier unlimited-at-quality framing. Two reads. (1) A capacity-gated cliff that slips a week at a time reads as a demand-signal Anthropic is still absorbing — the Fable 5 restoration week was Jul 1, and if the tier-1 buyer cohort still can't fit the credit meter by day 7, the economics of the metered regime are being re-priced in real time; the $10 / $50 credit meter is now a competitive number rather than a hostage number, which is why Grok 4.5 (item 05) and GPT-5.6 Terra / Luna (item 06) both land inside the same 48-hour window. (2) The "as capacity allows" language is the same clause Anthropic used yesterday to justify the TeraWulf $19B lease — 2H 2027 initial-capacity, early 2028 full-ops — which means the Fable 5 supply story is not "until Q4" but "until the Kentucky campus turns on," and the credit meter is the rationing instrument that buys the intervening 18 months.

The first legal-agent unicorn whose customer base is explicitly the buy-side of legal services — and the outcome-priced model is the story. Per the Bloomberg, TechCrunch, PRNewswire, SiliconANGLE, Artificial Lawyer and LawSites reads of the round: the $30T-AUM customer figure is what separates Norm from the Harvey / Ironclad law-firm-side cohort — Norm sells directly to in-house legal teams at asset managers, banks and insurers, which puts it on the same buyer-side as Taktile's agentic decision platform at regulated financial institutions (previously covered). Two reads. (1) A "charge by outcome" pricing model is the Agentforce Help Agent per-resolution pattern (previously covered Jul 1) applied to legal work — the frontier-agent commercial framing that ties buyer spend to a buyer-tracked outcome is now shipping in both customer service and legal counsel within a fortnight, which is the first time outcome pricing appears in two white-collar verticals in the same cycle. (2) A Norm Law AI-native firm that supervises its own agents is a vertically-integrated model that skips the law-firm middleman — the Fenwick LLP on the cap table is notable because Fenwick is the incumbent whose margin Norm is arbitraging, which is the same land-and-cannibalise pattern Salesforce Agentforce used against SI partners on the per-resolution launch.

The largest hyperscaler forward-deployment commitment in the industry — and the NANDA citation is the story. Per the Microsoft announcement, TechCrunch, CNBC, GeekWire, The Next Web and Yahoo Finance reads: MIT's Project NANDA "State of AI in Business 2025" found 95% of enterprise generative-AI pilots delivered zero measurable P&L impact — the most-cited enterprise-adoption statistic of the past year — and Microsoft citing it explicitly to justify a 6,000-person embedded consultancy is the first hyperscaler admission that software delivery alone doesn't move the P&L. Two reads. (1) A Palantir-shape (FDE) delivery model deployed at Microsoft-scale is the shape Google's Professional Services and AWS Professional Services pointed at last cycle but neither publicly staffed up to 6,000 engineers for — Palantir's FDE team is ∼800, and Microsoft just committed to 7.5× that headcount inside enterprise buyers specifically to close the NANDA gap. (2) The LSEG and Novo Nordisk early-customer names are worth reading: LSEG is the same OpenAI-Anthropic-Microsoft market-data buyer, and Novo Nordisk is the drug-discovery buyer that Anthropic's Claude Science is chasing — Microsoft is putting engineers on the ground at the same accounts where Anthropic and OpenAI are landing the agent layer, which resets the account-control contest.

The data-agent arms race gets its ERP-anchor closer — and the lakehouse-as-agent-foundation pattern is the story. Per the Futurum Group write-up: SAP is the buyer of first resort for ERP data at >77% of the Fortune 500, and pairing that first-party data with an Iceberg lakehouse (Dremio) plus an existing Databricks partnership plus Joule as the agent layer is the full stack Snowflake was assembling with Cortex Sense and Snowflake CoWork. Two reads. (1) An agentic BDC that queries across SAP + non-SAP without data movement is the zero-copy architecture the agent layer needed to escape the "pipeline your data first" tax — Dremio already ran that pattern on ∼3,500 customer instances, and folding those instances into SAP BDC lands SAP's agent surface on top of a real fleet the day the deal closes. (2) SAP extending rather than replacing its Databricks relationship reads as coexistence between Business Data Cloud and Databricks Data Intelligence Platform, which is the same polycentric data-layer bet enterprise buyers are increasingly demanding — and it gives SAP a real answer to Salesforce Data Cloud at the same Palexpo week Benioff was selling the $4B European Agentforce commitment (yesterday's item 05).

The European counterpart to the CAISI clearance (item 06) — and the "EU evaluation capacity" framing is the story. Per the European Commission release and the Digital Strategy write-up: the plan doesn't create new law; it points ENISA at secure access and a testing platform for the critical-sector cohort, which is the same cohort the Aug 2 EU AI Act fines-live deadline will start measuring against. Two reads. (1) A European evaluation capacity that lives at ENISA — and not at the AI Office inside DG CNECT — is a procedural choice: it puts cybersecurity rather than AI-safety at the centre of the capability assessment mandate, which is the same frontier-cyber framing the Five Eyes Jun 22 joint statement anchored, and the same CAISI pattern the Jun 2 EO set up in the US. (2) The Palexpo UN Global Dialogue closed Jul 7 (previously covered) — and the Commission publishing its Cybersecurity + AI plan the same day is a venue-parallel move: the UN is a state-only dialogue, the EU is a bloc-scale regulator, and the two layers now have coincident calendar entries the first time this cycle, which is the shape a Brussels-effect push takes when it starts moving from rule text to infrastructure.

The Bedrock AgentCore rate-limit envelope catches up to the enterprise fleet demand the GA at NY Summit surfaced — and the 8× InvokeAgentRuntime jump is the story. Per the AWS What's New bulletin: enterprise fleets running on AgentCore Harness have been running into 25 TPS-per-agent ceilings that made parallel fan-out patterns impractical, and the 200 TPS new ceiling brings AgentCore into the same operating envelope as the Anthropic message-batches API and OpenAI's enterprise-tier rate limits. Two reads. (1) A 500 → 5,000 concurrent sessions per account in the two flagship regions signals AWS is catching a parallel subagent workload shape at the same infrastructure layer where Claude Code's subagent-delegation fix just landed (previously covered Jul 7) and where Bedrock Agents Classic closes to new customers on Jul 30 — the fleet pattern is now the default assumption, not the edge case. (2) A 400 sessions per minute per endpoint container default is the shape a SaaS platform builds on top of — the independent agent-platform tier (Wayfound, Salesforce Agentforce, ServiceNow AI Agents) can now assume a Bedrock-hosted floor of ∼400/min without a formal quota-increase ticket, which lowers the time-to-launch curve for any agent-native startup whose first customer runs on AWS.