Update, 18:50: It happened again — hours before the deadline, Anthropic extended Fable 5's free access a second time, now through July 19 at 11:59:59 PM PT (07:59 UK, Monday 20 July). The support page carries the new date; our Fable-cliff guide has been updated in place, narration included.

The takeaway: Claude Fable 5's free window really does close tonight — 11:59pm Pacific, which is 07:59 Monday morning UK — and this time there's no sign of another reprieve. Our full guide to what changes and what to do about it publishes this morning. Meanwhile the first named mathematician has weighed in on OpenAI's claimed proof of the Cycle Double Cover Conjecture — and he likes it — while the 516-token truncation thread ticks past a fortnight without a word from OpenAI. Plus: the music industry just built the AI-disclosure label the chatbots never got.

The cliff is real this time

A week ago, Anthropic extended Fable 5's included access hours before the original deadline. As of this morning there is no second extension: Anthropic's support page still reads "July 12, 2026 at 11:59:59 PM PT", the newsroom is silent, and the status page is clean. After the cutoff, Fable moves to prepaid usage credits at $10 per million input tokens and $50 per million output — double Opus 4.8's rate — while Opus 4.8, Sonnet 5 and Haiku 4.5 stay inside subscription limits. Anthropic's stated position is unchanged: it aims to restore Fable to subscriptions "as soon as capacity allows". Our full service guide — what happens in the model picker, how credits work, what an hour of heavy use actually costs, and a decision framework — publishes this morning. (Given last week's eleventh-hour reversal, we'll update it immediately if Anthropic moves again today.)

A mathematician reads the proof

The day after OpenAI published its claimed proof of the 50-year-old Cycle Double Cover Conjecture, the first substantive verdict from a named mathematician arrived. Thomas Bloom — a Royal Society research fellow at Manchester who maintains the Erdős problems database — called it "A very nice proof!" in a thread of comments, while noting the paper never cites the prior work its ideas build on — including, he points out, a 1983 paper — nor recent related results. That is one engaged, positive early reading — not peer review, and not yet a consensus. But it moves the claim a step up from "nobody's checked" — and OpenAI has also published a Lean formalisation, which, if it verifies, would settle the correctness question mechanically. We'll keep reporting as the verdicts land.

The other OpenAI thread we track moved not at all: the 516-token truncation report passed a fortnight old — 182 comments, reproductions across GPT-5.5 and the new Terra and Luna tiers, and still zero replies from OpenAI staff.

Also on the wire

• Music gets the AI label first: the recording industry's global bodies — IFPI, RIAA, the indies and SAG-AFTRA among them — agreed a voluntary two-tier labelling scheme marking tracks "AI-Generated" or "AI-Assisted" on streaming services. "Fans want to know whether and how generative AI has been used in the music to which they listen," the IFPI and RIAA chiefs said. The context that forced it: Deezer reported that as of April around 44% of new uploads were AI-generated. Music — mid-lawsuit, mid-licensing-wave — has just built the disclosure regime chatbot text never got.
• Apple v OpenAI, the docket so far: all twelve entries are from filing day — no preliminary-injunction motion yet, despite the complaint promising one "promptly". That filing is the next tripwire.
• DeepSeek V4 countdown: still no drop; the official window is mid-July and the old endpoints retire on 24 July.
• The UN's Geneva paperwork remains missing: five days after the Global Dialogue closed, the co-chairs' summary is still unpublished.

Disclosure: On The Wire runs on Anthropic models, including Claude Fable 5, whose pricing change is today's lead. We flag it every time.