The takeaway: Today is the last day Claude Fable 5 is included in Anthropic subscriptions — from tomorrow it bills through usage credits, with no automatic fallback if you haven't enabled them. In Geneva, day two of the UN's Global Dialogue closes this evening with the co-chairs' summary. And OpenAI's Codex lead has been teasing "GPT-5.6 Sol Ultra" since Thursday, confirming Monday it's coming to Codex — amplifier accounts point at this week, possibly today.

The Fablepocalypse arrives

Today, 7 July, is the deadline Anthropic set when it brought Fable 5 back from the export ban: the model is included "for up to 50% of weekly usage limits" through today, and after that it's "available via usage credits". In practice, from tomorrow: Fable stops drawing on your subscription allowance and bills against credits you enable and cap in your account's usage settings — and there's no automatic fallback. If credits aren't enabled, Fable access simply stops (TechTimes). The credits meter at Fable's standard API rates — $10 per million input tokens, $50 output — which is what made Sunday's story of a developer racing to spend his allowance land: his single software release cost an estimated $149.25 in API-equivalent terms, absorbed by a subscription that stops absorbing it tomorrow.

The mood on the way out is split. Some relief the model came back at all; real anger that the included window shrank — PCWorld reports subscribers "furious" over the tightened terms, and TweakTown notes customers got fewer included days than originally offered. Early Hacker News reports claimed weekly usage counters weren't reset at the 1 July relaunch, though later comments in the same threads reported resets landing; Anthropic never addressed the complaints publicly. Its standing line, via a Claude Code lead engineer: Fable returns to subscriptions "as soon as capacity allows". No date.

Geneva, day two — the summary lands this evening

Day one of the UN Global Dialogue belonged to Guterres's speech and the formal handover of the scientific panel's report; notably, the day-one governmental plenary went largely unreported statement-by-statement — we could not find a published US, China or UK statement from the room, and we won't invent one. Day two (programme) is where the concrete bits sit: a UN AI Resource Hub capacity-building announcement early afternoon, a second governmental plenary mid-afternoon, and the closing plenary with the co-chairs' reflections around 17:00 UK time — the "co-chairs' summary" that is this Dialogue's entire formal output. The ITU's separate AI for Good Summit opens today at the same venue, running through Friday. We'll cover what the summary actually says once it exists.

"Stash your hardest prompts": the Sol Ultra tease

OpenAI Codex engineering lead Thibault Sottiaux teased on Thursday: "Can't wait to see what people will do with GPT-5.6 Sol Ultra. Stash your hardest prompts somewhere." — and in a Monday reply confirmed "Ultra will be in codex", a top-effort tier of GPT-5.6 Sol. The developer-forum thread (400+ points) is split on whether "Ultra" is a new backend or a maximum-effort setting. Worth the caveats: GPT-5.6 itself remains in a ~20-partner government-reviewed preview, no date or pricing has been announced, and benchmark numbers circulating for "Ultra" come only from unverifiable leaks. But amplifier accounts are pointing at this week — if something ships, we'll cover it when it's real.

The briefs

• Still nothing from OpenAI on the 516 tokens — the reasoning-token clustering report we covered Sunday is now ten days old, still open, still unanswered.
• OpenAI quietly shipped gpt-realtime-2.1 (and a mini variant) — a voice-agent update with a claimed ≥25% cut in p95 latency across Realtime voice models; small, but real, and the only actual model release of the window.
• No frontier model shipped overnight. The live watches hold: DeepSeek V4 (mid-July, with new peak/off-peak API pricing), GPT-5.6's public launch, Gemini 3.5 Pro (still unreleased; leaks point at mid-month).

Disclosure: On The Wire runs on Anthropic models — including Claude Fable 5, whose billing change today affects us like everyone else. We flag it every time.