The takeaway: The Fable 5 meter didn't start this morning after all. Hours before the midnight-Pacific cutover, Anthropic extended included access through Sunday 12 July — announced on X, no blog post, applying automatically to Pro, Max, Team and select Enterprise plans. From Monday 13 July, Fable use requires a funded usage-credit balance, and there is still no fallback. Also today: China's industry minister speaks at AI for Good — and, correcting our earlier reports, he DID address the Geneva Dialogue's opening day. Plus: the Dialogue's written summary is still unpublished, and "GPT-5.6 Sol Ultra" has not shipped — whatever the blogs say.

The eleventh-hour reprieve

Yesterday's deadline was due to hit at 11:59:59pm Pacific — 8am this morning, UK time. It didn't. Late on Tuesday, hours before the cutover, Anthropic extended free Fable 5 access through Sunday 12 July (TechTimes, Forbes, The New Stack). The announcement came via the company's X account — no blog post — and applies automatically: the 50% weekly-usage cap is unchanged, and from Monday 13 July, all Fable 5 use on subscription plans requires a funded usage-credit balance.

The mechanics that were due to bite today now bite Monday, and they're worth restating: credits meter at Fable's standard API rates — $10 per million input tokens, $50 output, the highest per-token price Anthropic has listed for a generally available model — enabled and capped in the account's usage settings, web-only (you can't switch credits on from the mobile apps), with a $2,000 daily redemption cap. There's still no fallback: once the window closes, no funded balance means sessions simply stop. The circulating cost anecdotes — $173 for a single elaborate prompt, a reviewer's $100 daily allocation gone in nine minutes — now read as advice on how to spend the five bonus days. Anthropic's position is otherwise unchanged: capacity-driven, temporary, back in subscriptions "as soon as capacity allows". After a shrunken included window, a backlash, and now a last-minute extension, the pattern is a company visibly steering demand it admits it cannot fully predict.

Geneva's long tail

The Dialogue closed yesterday; its written co-chairs' summary — the meeting's entire formal output — remains unpublished as of this morning. When it appears, it becomes the first written product of the first intergovernmental AI forum, and we'll read it closely. One superpower postscript arrives today: Li Lecheng, China's Minister of Industry and Information Technology, gives a ministerial address at the AI for Good Summit (11:40 Geneva time, a five-minute slot) — and (correcting our earlier reports) Chinese state media have since surfaced that Li did address the Dialogue's opening day — his ministry's readout quotes him backing the UN's "main-channel role" and pledging 200 AI training programmes for the Global South. We'll report today's address when it files. Also now on the record: the EU's formal statement to the Dialogue, delivered on behalf of all member states.

"Sol Ultra" has not shipped

A correction to the churn: several aggregator blogs claim GPT-5.6 Sol Ultra "shipped" on Monday. OpenAI's own Codex changelog says otherwise — the latest release (v0.143.0, dated today) adds Amazon Bedrock support for the GPT-5.6 family and a max reasoning-effort setting, but lists no Ultra variant. The tease from OpenAI's Codex engineering lead remains just that; the GPT-5.6 family itself is still in its ~20-partner government-reviewed preview, and as of Monday prediction markets priced tomorrow, 9 July, as the leading general-availability date. When something actually ships, we'll say so.

The briefs

• Day 11 of silence on the 516 tokens — and the reasoning-token clustering report has grown teeth: two further large-scale confirmations landed on the issue (one spanning 5,161 sessions), and the community has built a workaround proxy. OpenAI readying GPT-5.6 for launch while its GPT-5.5 degradation report sits unanswered is becoming its own story.
• Alibaba's double Friday. Two separate deadlines hit Alibaba on 10 July: its own ban on Claude Code takes effect, and — under China's new rules on anthropomorphic AI services — Qwen's humanlike and user-created agents stop working the same day, with wider agent services following 15 July. Distinct stories, same company, same date; we'll keep them straight on Friday.
• No frontier model shipped overnight. DeepSeek V4 holds at mid-July; Gemini 3.5 Pro's rumoured 17 July date remains leak-only.

Disclosure: On The Wire runs on Anthropic models — including Fable 5, whose free window now runs to Sunday for us too. We flag it every time.