The takeaway: GPT-5.6 launches publicly today — not live as we publish, with the switch expected to flip mid-afternoon UK time. The pricing is locked, the government gate is lifted, and the early testers are split. Meanwhile a genuinely fresh piece of British science: the Met Office and the Alan Turing Institute say they've cracked one of AI weather forecasting's biggest problems — forecasts that look right and obey physics. Plus: tomorrow is Alibaba's double deadline, and the 516-token question enters day twelve.

Launch day — not live at press time

Today is the day OpenAI said GPT-5.6 goes public — all three variants, globally, after Washington lifted the preview gate. As we publish, nothing has flipped yet: no launch post, no model-picker sightings, the preview page still the newest official word. Expect the switch mid-afternoon UK time, on the US morning; we'll cover the launch when it's real.

What's already firm: pricing holds at Sol $5/$30 per million tokens, Terra $2.50/$15, Luna $1/$6, with Sol on Cerebras hardware promised at up to 750 tokens per second this month. "Sol Ultra" — the teased multi-agent Codex tier — appears in reporting but on no official OpenAI page; today's launch post is where that gets settled. And the early testers disagree in the way that matters: preview-cohort voices call it "the best model I've ever used" and "world leading in computer use", while dissenter Matt Shumer found that "for almost every task I tested, Fable was quite a bit better" (Axios's collation). Independent benchmarks start today.

A physics lesson for AI weather

Fresh from the Met Office and the Alan Turing Institute: peer-reviewed research showing their jointly built FastNet model can produce forecasts that are physically realistic, not just statistically accurate. The problem it addresses is the known dirty secret of AI weather models — they score well on averaged error metrics while producing blurred, physically unrealistic fields underneath. The fix is a modified training objective (a spherical-harmonic loss function, for the technically inclined) that penalises unphysical output — and the result matches the Met Office's operational Global Model on headline accuracy while staying physically coherent. It's not operational yet, but it's British, primary-sourced, published yesterday, and a genuine step in the argument about whether AI forecasting can be trusted rather than merely impressive.

The briefs

• Tomorrow is Alibaba's double Friday — both legs now verified. The Claude Code ban takes effect, and separately — ahead of China's new rules on humanlike AI services, which take effect 15 July — Qwen's anthropomorphic and user-created agents shut down the same day. Full coverage tomorrow.
• Day twelve on the 516 tokens. Still no OpenAI response — a silence worth re-checking the moment today's launch lands.
• Geneva's paperwork is still pending: the co-chairs' written summary remains unpublished two days after the Dialogue closed. One precision note on our China coverage: Minister Li Lecheng's substantive Geneva remarks came at Monday's Dialogue opening (now covered in English — open-source AI as "a shared asset for all humanity"); his Wednesday AI-for-Good slot was scheduled for five minutes and remains unreported. China's next stage is its own: the World AI Conference in Shanghai, 17–20 July.

Disclosure: On The Wire runs on Anthropic models. We flag it every time we cover the company or its competitors.