1. Washington reopens the gate

On June 30 the Commerce Department lifted the export controls it imposed in June, and Anthropic restored Fable 5 globally on July 1 across Claude.ai, the API, Claude Code, and Cowork. The redeployment ships with a new government-coordinated safety classifier that blocks the flagged technique in over 99% of cases and reroutes those requests to Opus 4.8. Last issue this was the lead story going the other way. Nineteen days later the same lever moved in reverse.

  • The point is not that Fable is back. It is that frontier access is now a dial Washington can turn both ways, on its own timeline.
  • The open question users are watching: whether the new guardrails cost everyday performance, or only block the narrow flagged category.

2. OpenAI offers the government equity

OpenAI has reportedly proposed giving the US government a 5% stake, and suggested other leading labs contribute similar stakes into a public fund modeled on the Alaska Permanent Fund. Talks are early and conceptual, and any deal would likely need an act of Congress. Read one way it is a public dividend on AI; read another, it is insurance, when the same government can gate your model and shape your IPO.

  • If you sit outside the US, the question is whose incentives now sit behind your critical AI vendor.

3. Near-frontier AI hits the free tier

Anthropic shipped Claude Sonnet 5 on June 30, near-Opus performance at a fraction of the cost, and made it the default for every free and paid user. Introductory API pricing is $2 and $10 per million tokens through August 31. Last issue's story was enterprises capping AI spend and routing simple jobs to cheaper models. This is the supply side of the same shift: the frontier just got cheap enough that routing is the obvious move.

  • The strategic question flips from "which model is best" to "which model is good enough for this job, at this price."

4. Meta starts reselling its compute

Meta is building a cloud business, reportedly called Meta Compute, to sell access to the AI infrastructure it has been buying at a $115-145B annual pace. The stock rose about 8%; CoreWeave and Nebius, which hold $21B and up to $27B in contracts to supply Meta, fell double digits on the read that a customer just became a competitor. It follows SpaceX, which is already leasing capacity to Google and Anthropic.

  • When the biggest spenders start reselling the overflow, compute starts behaving less like a moat and more like a commodity.

5. Mystery model was actually Chinese

The coding model that topped OpenRouter charts for two months as the anonymous "Owl Alpha" turned out to be LongCat-2.0, from Chinese delivery giant Meituan, revealed June 30. It is a 1.6-trillion-parameter model priced at roughly a tenth of the US flagships, and Meituan says it was trained end-to-end on domestic chips with no Nvidia. The benchmark claims are self-reported and the open weights are not yet downloadable, so for now use runs through Meituan's servers.

  • A near-frontier model can now win global adoption before anyone knows who built it or on what hardware. Price and provenance both moved this week.

More AI stories of the week

  • Robert Maciejko: The Board's AI Opportunity. Only 8% of directors rate themselves strong on AI, yet they are the people governing it; the argument is that the gap is an opportunity if boards create a safe space to learn first, and that the board's role is power steering, not the brake.
  • South Korea commits $880B to chips and AI. President Lee announced roughly 1,350 trillion won in mostly-private investment led by Samsung and SK Hynix, largely for new fabs and data centers, one of the largest coordinated national tech bets on record.
  • Palantir's Karp torches the token model. On CNBC, the Palantir CEO argued OpenAI's and Anthropic's token pricing is oversold and enterprises should own their data and compute; opinion, but a notable one given his Nvidia open-model tie-up.
  • Heavy AI adopters are hiring more, not less. A Ramp/Revelio study of 21,000+ US firms found companies investing most in AI grew headcount about 10% over two years, with entry-level roles up 12%, cutting against the jobs-apocalypse narrative (vendor data, but firm-level and empirical).
  • Ford rehires veteran engineers after AI fell short. The automaker brought back 350 "gray beard" specialists after automated quality systems disappointed, crediting them with hundreds of millions in warranty and recall savings; a concrete case of AI oversold in one operational setting.
  • Why vertical AI founders are leaving the model labs. A founder who sold his last company for $780M argues the frontier labs are becoming competitors, not vendors, and that proprietary data and evals, not the model, are the real moat.
  • China's grey market for Claude tokens. An Oxford researcher's analysis (opinion) of the "transfer station" economy that resells Anthropic access at about 10% of list price, arguing the bigger governance problem is not distillation but eroded provider traceability, as every new control spawns a matching evasion layer.
  • Ed Zitron makes the AI bear case on CNBC. A prominent AI skeptic (opinion) laid out doubts on AI ROI and OpenAI's finances on mainstream business TV; the airing itself is notable given how rarely such views run there.
  • "Europe 2031": a scenario of AI-driven decline. A five-year forward scenario (opinion) that runs from the real Fable/Mythos export-control story into a future where Europe, dependent and outbuilt on compute, loses the ability to shape its own future; a sharp articulation of the access-and-dependency question.
  • Anthropic launches Claude Science. An AI workbench that pulls scientists' fragmented tools into one environment, with 60+ scientific databases, reproducible artifacts traced to their code, and on-demand compute, now in beta; Anthropic also said it will run its own drug-discovery programs for neglected diseases.
  • California signs a first-of-its-kind state AI deal. Every state agency plus opting-in cities and counties get Claude at a 50% discount, with free workforce training.
  • The UN opens its first global AI governance dialogue. The inaugural Geneva session runs July 6-7, informed by a preliminary report from a new UN scientific panel co-chaired by Yoshua Bengio and Maria Ressa.