The week's real product was a precedent: access to a frontier model is now something a government grants.
Claude Fable 5 came back on 1 July, restored across the API, Claude.ai, Claude Code and Cowork after a 19-day worldwide shutdown under export controls. Mythos 5, the unrestricted sibling, returns only to roughly 100 vetted US companies and federal agencies after government review; Commerce lifted the order on 30 June saying "appropriate safeguards are in place", and restored access ran at half the normal usage limits until 7 July. A frontier model was switched off by the state and switched back on by the state, inside three weeks. CNBC
OpenAI ran the same gauntlet from the other direction. GPT-5.6 (a three-tier family: Sol, Terra, Luna) was previewed on 26 June but gated to around 20 government-vetted partners, shared with the US government before release, general availability promised "in the coming weeks". Two labs, two gated releases; pre-release government review became practice this week, before any framework formally requires it. OpenAI
The Pentagon lost round one
Judge Rita Lin granted a preliminary injunction blocking the Pentagon from enforcing its "supply chain risk" designation on Anthropic, and the unsealed emails were ugly for the government. Under Secretary Emil Michael told Dario Amodei the two sides were "very close" on contract terms the day after the risk label was finalised, and the underlying memo cited Anthropic's "hostile" press posture, which the judge read as classic First Amendment retaliation. The actual dispute underneath: Anthropic won't drop its bans on fully autonomous weapons and domestic mass surveillance. This is the first court test of whether a lab can hold usage-policy red lines against government pressure, and the outcome sets the weight every vendor's acceptable-use commitments carry. Gizmodo · Tech Brew
And OpenAI offered Washington equity
What started the week as an unverified FT line firmed up into confirmed reporting: OpenAI has proposed the US government take a 5% stake (about $42.6B at its $852B valuation), with Sam Altman pushing for every leading US lab to pay into an Alaska-style "Public Wealth Fund". Talks are early and it may need an act of Congress. Stack it on the export saga and the Pentagon fight, and state entanglement with the labs is the defining industry dynamic of the summer. CNBC
The buildout didn't pause for any of it
Alphabet closed a record $84.75B equity raise for AI compute, the largest corporate equity financing ever, with $10B of it a private placement to Berkshire Hathaway. OpenAI announced its Jalapeño inference chip; Anthropic is in early talks with Samsung on a custom chip using the 2nm process. Microsoft launched "Frontier Company", a $2.5B operating business with 6,000 experts dedicated to making enterprise AI deployments actually work, two days after AWS committed $1B to its own forward-deployed engineering venture. Qualcomm bought Modular for $3.9B. Half-year venture numbers landed too: a record $510B globally, more than all of 2025, with OpenAI and Anthropic taking $217B of it between them. TechCrunch · Crunchbase
China owns the open layer
Chinese open-weight models hit roughly 61% of tokens consumed on OpenRouter; four of the top five most-used models are Chinese, and Llama has fallen off the list entirely. Meituan's LongCat-2.0 (a 1.6T-parameter MoE trained end to end on around 50k domestic chips) sharpens the argument that US gatekeeping is handing China the open-source edge. The contest is no longer best single model; it's who supplies the tokens the world runs on. Anthropic, meanwhile, closed the offshore workarounds Chinese firms were using to reach Claude (Ant issuing staff accounts via a Singapore subsidiary; ByteDance reimbursing VPN-bought subscriptions) and disclosed a 25,000-account distillation attack to senators. DataGravity · BanklessTimes
Quietly useful
Claude Sonnet 5 shipped on 30 June and is now the default for Free and Pro users: the most agentic Sonnet yet, near Opus 4.8 on many tasks, with intro API pricing of $2/$10 per million tokens until the end of August (then $3/$15). A materially better default at a lower price resets the cost basis of anything built on the mid tier. And token cost governance went mainstream as an ops concern: Tesla capped employee AI spend at $200 a week (xAI products exempt, pointedly), Uber at $1,500 a month after blowing its 2026 AI budget by April. Anthropic · Electrek
What to watch
Governance week. The UN's first Global Dialogue on AI Governance opens in Geneva on 6 July, and the White House voluntary frontier-release framework could land the same week; if both produce text, the compliance baseline for enterprise AI gets written in five days. The Pentagon injunction is the deeper story though. It decides whether a lab's red lines survive contact with its largest possible customer.