The US lifts its export controls and Anthropic’s top models come back online

The Commerce Department lifted the export controls it imposed on June 12, and Anthropic restored global access to Fable 5 on July 1 across Claude.ai, the Claude Platform, Claude Code and Claude Cowork. The restriction was the first the US applied to an AI model rather than to hardware, and it kept the company’s most capable models offline for close to three weeks. Mythos 5 remains limited to vetted Project Glasswing partners.

Anthropic used the moment to press for a durable pre-release process applied equally to all frontier developers, and it is working with Amazon, Google and Microsoft on a shared way to score how dangerous a given jailbreak is.

For buyers: vendor access to frontier models can now change on short notice for policy reasons, a new variable in procurement and continuity planning.

The White House moves toward voluntary standards for frontier-model releases

The Financial Times reported the administration is in advanced talks with OpenAI, Google and Anthropic to set voluntary standards for how new frontier models are released, covering benchmarks, testing timelines and access rules. An announcement could come as soon as the following week, and the administration faces an Aug. 1 deadline to define how it assesses security risk in new models. This would formalize the kind of government review that produced the Fable 5 suspension and the GPT-5.6 delay.

For buyers: expect model launch timing to become less predictable, and build flexibility into roadmaps that depend on new capabilities.

OpenAI floats giving Washington a 5 percent stake in the company

The Financial Times reported that OpenAI proposed giving the US government a stake of about 5%, worth roughly $42.6 billion at its $852 billion valuation, and suggested peers such as Anthropic, Google and Meta contribute similar equity to an Alaska-style public fund. Talks are early and could require an act of Congress. The report followed OpenAI’s decision days earlier to delay the full public launch of GPT-5.6 at the government’s request.

For US and UK executives: once Washington holds equity in the labs, buyers abroad may re-examine neutrality and data-sovereignty assumptions about American providers.

Court unseals the Anthropic-Pentagon emails behind the blacklisting

Court documents unsealed July 2 in the Northern District of California, first reported by the Wall Street Journal, made public the exchange between Anthropic CEO Dario Amodei and Pentagon official Emil Michael over the Defense Department’s demand for access to Claude for “all lawful uses.” Anthropic held its line against fully autonomous weapons and domestic mass surveillance, which led the Pentagon to label the company a supply-chain risk.

Separately, financial disclosures show Michael held stock in xAI, an Anthropic rival, a conflict reported by The Guardian and ProPublica, and the Pentagon has since moved to bring xAI’s Grok onto classified systems.

For buyers: the case tests whether an AI vendor can enforce use limits on a government customer, which shapes how any enterprise writes usage terms into AI contracts.

UN and ITU launch the AI for Good Global Commission

The United Nations and its International Telecommunication Union launched the AI for Good Global Commission on July 2, the first UN-level body to seat the leaders of major AI companies alongside heads of state. Salesforce CEO Marc Benioff and Rwandan President Paul Kagame co-chair it, with members including Nvidia’s Jensen Huang, Amazon’s Andy Jassy, Anthropic co-founder Jack Clark, Cohere’s Aidan Gomez and Microsoft’s Brad Smith. Its first meeting is July 8 in Geneva, during a week that also includes the UN Global Dialogue on AI Governance and the ITU AI for Good summit.

For buyers: the commission is unlikely to produce a treaty, but the terminology and voluntary standards it sets tend to precede binding rules, so multinationals should track it.

Meta bars its own engineers from Claude Code and Codex

The Information reported that Meta’s Applied AI division told engineers to stop using Anthropic’s Claude Code and OpenAI’s Codex without approval, with one internal memo directing some teams to pause tasks that relied on the tools. The concern is inadvertent distillation, where outputs from rival models could enter Meta’s training data or benchmarks and create breach-of-terms exposure while Meta builds its own coding tool, MetaCode. It is one of the first public cases of a major lab restricting staff use of a competitor’s coding agent on distillation grounds.

For buyers: data-exposure and IP questions around AI coding tools now reach the largest and most technically sophisticated buyers, not only regulated firms.

Alibaba bans Claude Code as the distillation fight with Anthropic escalates

Alibaba issued an internal notice on July 3 banning Claude Code from July 10 and directing staff to its own Qoder tool, after researchers flagged that recent Claude Code versions could detect users tied to Chinese AI labs. The move follows Anthropic’s late-June accusation that Alibaba’s Qwen lab used roughly 25,000 fraudulent accounts to run about 28.8 million conversations and extract knowledge from Claude, along with Anthropic’s broader push to block Chinese firms that reach its models through overseas subsidiaries.

For buyers: model-level intellectual property is becoming a trade flashpoint alongside chips, which matters for any enterprise weighing vendor lock-in and cross-border access.

Meituan open-sources LongCat-2.0, trained end-to-end on Chinese chips

Meituan released LongCat-2.0 on June 30, a 1.6 trillion-parameter mixture-of-experts model built for agentic coding, with a 1-million-token context window and an MIT license that permits commercial use. Meituan says it is the first trillion-parameter model both trained and served entirely on domestic Chinese ASICs, with no Nvidia GPUs, and it spent two months near the top of OpenRouter developer charts under the codename Owl Alpha. The company’s own benchmarks put it slightly ahead of GPT-5.5 on SWE-bench Pro, though full weights and independent testing are still pending.

For buyers: it adds a lower-cost, US-independent option for coding workloads and shows export controls are not stopping frontier-scale Chinese training.

Anthropic ships Claude Sonnet 5 as the new default

Anthropic launched Claude Sonnet 5 on June 30 as the new default for Free and Pro users. The company positions it as its most capable agentic Sonnet, performing close to Opus 4.8 on many tasks while costing less than the prior Sonnet at introductory pricing through Aug. 31. The pitch targets a real complaint from Q2, when agent workloads ran up unexpectedly large token bills.

For buyers: teams running or planning agentic workflows should re-test cost and performance against this release before committing to a model.

California signs a statewide Claude deal the federal government would not

Gov. Gavin Newsom signed an agreement giving California state agencies, cities and counties access to Claude at a 50% discount. The state already runs Claude across several programs, including an internal assistant called Poppy piloted with more than 2,800 employees across 67 departments, plus uses at the DMV and in Medicaid casework. The deal stands out because the federal government has separately labeled Anthropic a supply-chain risk.

For buyers: state and federal postures on the same vendor now diverge, giving public-sector and regulated buyers a reference for large-scale discounted deployment.

Microsoft reportedly plans layoffs while AI spending climbs

Business Insider reported that Microsoft is preparing to cut fewer than 2.5% of its roughly 220,000 employees, up to about 5,500 people, with sales, consulting and Xbox expected to absorb most of the reductions. Microsoft had not confirmed the cuts, which would coincide with the start of its fiscal year and follow heavy AI infrastructure spending. The pairing of record AI investment with continued headcount reduction is now a pattern across Meta, Amazon and Microsoft.

For buyers: read this as AI capital spending being funded in part through cuts to customer-facing and support functions.