The takeaway: Apple sued OpenAI on Friday, alleging trade-secret theft in the hiring of its hardware engineers — the biggest new front in AI litigation since the copyright cases, and aimed squarely at OpenAI's device ambitions. On the same day, OpenAI claimed its top model setting had proved a 50-year-old mathematics conjecture in under an hour — a claim nobody has independently validated yet — while the 516-token truncation thread reached 183 comments and day 14 with still not one reply from OpenAI staff. And a service note for our readers: Claude Fable 5's free window ends tomorrow night, US Pacific time — early Monday morning UK.

Apple v OpenAI

Filed Friday in the Northern District of California, Apple's suit names OpenAI and its io Products hardware arm, plus two former Apple employees — Tang Yew Tan and Chang Liu — and alleges a pattern behind OpenAI's device push: recruiting Apple staff using code names, candidates reportedly asked to bring Apple hardware to interviews, and an unreturned laptop allegedly used to download confidential documents. The filing notes more than 400 former Apple employees now work at OpenAI, and asks for a preliminary injunction. OpenAI's io — the Jony Ive venture it absorbed — is the vehicle for its long-trailed AI device. Apple going to court over the talent pipeline behind it turns a rivalry into discovery, subpoenas and depositions. We'll cover the filing in detail today.

A proof nobody's checked — and a silence everyone can see

OpenAI's flashiest capability claim since launch arrived Friday: that GPT-5.6 Sol Ultra — the multi-agent setting — proved the Cycle Double Cover Conjecture, an open problem in graph theory since the 1970s, in under an hour using 64 coordinated subagents, with a proof PDF published on OpenAI's CDN. If it survives peer scrutiny it is a genuinely historic result. It has not yet been validated by anyone outside OpenAI, and mathematics has seen machine-assisted "proofs" collapse under review before — so it is a claim, not yet a result. We'll report what mathematicians find, whichever way it goes.

The same company's other thread tells a different story. The 516-token truncation report reached day 14 and 183 comments — with zero replies from OpenAI staff, a silence that has now outlasted a flagship launch. And the reports keep hardening: one tester ran GPT-5.6 Sol across 125 runs and found high accuracy but truncation-continuation rates of 52% at Extra High and 56% at the maximum effort setting; the cheaper Terra and Luna tiers continue to show wrong answers on tiny, clustered reasoning budgets. The juxtaposition writes itself: a company publishing a 64-agent proof of a 50-year conjecture while declining, for a fourteenth day, to acknowledge a reproducible report that its shipping models truncate their own thinking.

Also on the wire

• Fable cliff, tomorrow night: Claude Fable 5's free window ends Sunday 11:59pm Pacific — 07:59 Monday morning UK. After that it moves to prepaid usage credits at $10/$50 per million tokens. If you've been using Fable on a subscription, Sunday is the day to check your settings; we'll publish a fuller guide tomorrow.
• The Geneva paperwork is still missing: four days after the UN's Global Dialogue closed, its co-chairs' summary remains unpublished — while the UN's own first independent AI assessment, co-chaired by Yoshua Bengio and Maria Ressa, warns that risks are outpacing governance. The governments' verdict on themselves is the missing document.
• DeepSeek V4 countdown: the formal release window is mid-July, and the old endpoints retire on 24 July — two weeks, one migration.
• Tencent's Hy3 is still absent from Artificial Analysis's independent leaderboards two days after its vendor-claimed near-frontier scores; the verification gap stands.
• Alibaba's Friday deadlines passed without fresh reporting on how the Claude Code uninstall and Qwen shutdown actually landed inside China — we're watching the weekend filings.

Disclosure: On The Wire runs on Anthropic models, including Claude Fable 5, whose free window features in today's update. We flag it every time.