Metered and Verified: Fable 5's Durable Access Terms Take Effect This Week
The post-restoration grace period ends now. When Anthropic redeployed Fable 5 on 1 July after the 19-day export-control suspension, subscribers got the model back inside 50% of their weekly plan limits as a capacity-management measure. That window closes this week: per multiple trackers, Fable 5 moves on 7 July to metered usage credits billed outside the standard plan at $10 per million input tokens and $50 per million output — double Opus 4.8's $5/$25, and Anthropic's most expensive pricing since the Opus 4.5 price cut (earlier Opus models listed at $15/$75). From 8 July, consumer plans (Free/Pro/Max) must also clear identity verification through third-party provider Persona — a government-issued photo ID plus a live selfie — before credits apply. App strings spell it out: "Your credits will be added once your identity is verified." API and Enterprise customers are exempt, for now.
This is not a temporary safeguard; it is the durable access model for Anthropic's most capable tier, and it makes concrete what June's restoration terms implied — the vendor has internalised the state's gating machinery and now operates it at the individual level. For European buyers there are two immediate questions. First, jurisdiction: access to the strongest commercial model is now conditioned on identity checks whose rules are set in Washington and enforced by a US verification provider — a CIO analysis notes the new privacy policy was shaped around exactly this regime. Second, data protection: a live-selfie flow means facial-geometry data from EU consumers processed by a US third party, a GDPR Article 9 question that DPAs will not leave unexamined. The enterprise exemption softens the blow today, but the machinery, once built, tends to travel.
A Week On, China's Biggest "Open" Model Still Has No Weights
Six days after Meituan announced LongCat-2.0 as open source under MIT licence — 1.6 trillion parameters, ~48B active, 1M context, trained entirely on Chinese ASIC superpods — the Hugging Face repository still carries a single banner where the checkpoint should be: "Model weights coming soon — stay tuned!" No weight files, no inference-provider deployment, downloads not even tracked. Meanwhile a second wave of aggregator coverage this weekend asserted the weights "have since been made available." They have not. We checked the repository this morning; it is the third time in a week that secondary coverage has claimed downloadable weights for this specific model while the repo says otherwise.
The gap matters more than the benchmark numbers. An MIT licence on an announcement page confers nothing if the artifact never ships: until the checkpoint is in your hands, LongCat-2.0 is a cheap hosted API in someone else's jurisdiction — exactly the caveat that applied at launch. The contrast with Europe's recent releases is instructive: Portugal's Amália shipped model, dataset and code under Apache 2.0 on day one; MiniMax M3's weights followed its API by about ten days. For European teams evaluating "open" models for sovereign deployment, the procurement test is simple and worth writing down: verify the artifact, not the licence headline. If the LongCat checkpoint lands this week, the sovereignty calculus genuinely changes — we are watching the repo, not the press releases.
193 States, One Table: The Access Question Goes Multilateral in Geneva
The inaugural UN Global Dialogue on AI Governance opens today at Geneva's Palexpo (6–7 July), co-chaired by El Salvador's Egriselda López and Estonia's Rein Tammsaar — the first platform convened by the UN General Assembly where all 193 member states, industry, civil society and the technical community formally share a table on AI. It anchors a packed Geneva week: the ITU AI for Good Summit follows 7–10 July with some 11,000 participants from 169 countries, and the new AI for Good Global Commission — Benioff and Kagame co-chairing, with Huang, Jassy, Brad Smith, Jack Clark and Aidan Gomez among members — holds its first meeting on 8 July. On the formal agenda: "bridging AI divides" — capacity, access and compute for the countries that have none of the three.
The timing gives the agenda its edge. This dialogue convenes after a month in which access to frontier models became an explicitly unilateral instrument — vetted-partner gating for GPT-5.6, a 19-day export-control suspension and conditional restoration for Fable 5, and a voluntary pre-release framework due from Washington by 1 August. For the states on the receiving end of those decisions, Geneva is the first venue with standing to ask whether "who gets access, on whose terms" should be negotiated anywhere other than one capital. The test of the next 48 hours is whether anything concrete emerges — export-control interoperability language, access commitments, anything beyond a communiqué — before the caravan disperses until the second session in New York in May 2027. Watch this space tomorrow.
Quick Hits
- The rule-writer becomes the rule-taker — Dean Ball, co-author of the White House AI Action Plan and until recently OSTP's senior AI policy adviser, starts today as OpenAI's Head of Strategic Futures, reporting to Jason Kwon — three weeks from writing frontier-access policy to advising the lab most shaped by it (citybiz).
- EU AI Act: four weeks to enforcement — The Commission's GPAI penalty powers (up to €15M/3% of turnover) activate 2 August; the Digital Omnibus that defers high-risk obligations to Dec 2027/Aug 2028 cleared its final Council vote 29 June and is expected in the Official Journal this month — publication would settle the timeline for good (artificialintelligenceact.eu).
- The July model window — Gemini 3.5 Pro is cleared for general availability this month (2M-token context, still in Vertex preview with no firm date), while GPT-5.6 remains gated to ~20 government-vetted partners with GA consensus around mid-July — both launches now shadowed by the voluntary release framework expected from Washington as soon as this week.