GPT-6
OpenAI shipped GPT-6 Astra on September 3 — to enterprises in the Trusted Access Program first, with the API and the paid ChatGPT plans promised “in the coming days” at launch. On the API side that window has since closed: our own hourly sweep read the plain id in the OpenRouter catalog at OpenAI’s base sticker on September 4 at 21:16 UTC — openai/gpt-6-astra, $10/$50 per million, 1,050,000 context — so Astra has been on the public API since Sep 4. The model id is gpt-6-astra. The sticker is $10 / $50 per million with $1 cached input — identical to Claude Fable 5.1 — and the context window is 1.05M tokens (922K max input) with 128K output. Artificial Analysis puts it at 67 on the Coding Agent Index, level with Opus 5 and Fable 5, and 61 on the Intelligence Index, equal to GPT-5.6 Sol and five points under Fable 5.1.
Leaks & news, newest first.
Astra is answering on the public API — our sweep read it at OpenAI’s base sticker
The launch-day “in the coming days” is settled on the API side, and our own hourly catalog sweep is what settled it. At 2026-09-04 21:16 UTC the sweep found the plain, non-variant id openai/gpt-6-astra in the OpenRouter catalog at $10.00 in / $50.00 out per million with a 1,050,000-token context — OpenAI’s own base sticker — alongside openai/gpt-6-astra-pro at the same price. Three third-party listings followed within eleven hours: vercel/openai/gpt-6-astra-fast ($20/$100, Sep 4 23:45), venice/openai-gpt-6-astra and venice/openai-gpt-6-astra-pro (both $12.5/$62.5, Sep 4 23:45 and Sep 5 08:06). A catalog listing is a primary source for exactly one fact — that this id is being served, at this price, with this context window — and that is precisely the fact the launch-day promise was about. What we do not claim: which paid ChatGPT tiers have it. This sweep reads catalogs, not OpenAI’s consumer rollout. via our own hourly sweep — OpenRouter · models.dev
GPT-6 Astra is out — trusted access first, public API within days
OpenAI released GPT-6 Astra: “Our most capable model, built for the hardest end-to-end work”, per its API docs, for complex reasoning, coding, computer use, research and document creation. Model id gpt-6-astra; $10 in / $50 out per million with $1 cached input; 1,050,000-token context, 128,000 output; knowledge cutoff April 30, 2026. It is rolling out to enterprises in the Trusted Access Program today, with the API and paid ChatGPT plans to follow. Artificial Analysis’ launch-day read: 67 on the Coding Agent Index (level with Opus 5 and Fable 5; Fable 5.1 leads at 70) and 61 on the Intelligence Index (equal to GPT-5.6 Sol). via OpenAI developer docs · Artificial Analysis
New leak: August target, 1.5M+ context, a new pre-train foundation
A fresh leak — circulating under both “GPT-5.7” and “GPT-6” labels, which itself tells you how firm it is — points at an August release: a 1.5M+ token context window, an entirely new pre-training foundation beyond the old base (with 10T-scale rumors attached), and substantial agentic/reasoning gains aimed at the Mythos/Fable tier. Nothing from OpenAI itself; if it ships as 5.7, the GPT-6 market doesn’t resolve. via WinCentral · Times of AI
GPT-5.6 goes public — and it is not GPT-6
OpenAI shipped its “most capable system yet” in three tiers — Sol, Terra, Luna — as a 5.x point release. The numbering now means the digit marks the generation and the tier names persist across it, so the next numbered drop is the one this page watches. Polymarket’s rules agree: 5.x releases don’t resolve the GPT-6 market. via OpenAI · CNBC
Altman’s “extra goblins” GPT-6 post — a joke, read as a hint
Altman posted a joke prompt reading “start training GPT-6, you can have the whole cluster. extra goblins.” Widely read as a playful nod that GPT-6 is next — not an official signal, and we don’t treat it as one. via Futurism · TechRadar
GPT-5.5 ships — the model many expected to be GPT-6
Instead of a numbered bump, OpenAI released the model reportedly codenamed “Spud” as GPT-5.5 — and it absorbed much of the memory and agentic direction previously teased for GPT-6, resetting expectations for the numbered successor to H2 2026. via lifearchitect.ai (reported) · FinancialContent
Altman’s clearest GPT-6 framing: memory first, coming faster
On the record with CNBC: GPT-6 will arrive faster than the gap between GPT-4 and GPT-5, and personalization is the point — “People want memory.” Still the most concrete official statement about the model. via CNBC
The official word.
“Our most capable model, built for the hardest end-to-end work.”
OpenAI, GPT-6 Astra model documentation · Sep 3, 2026
Hear it the second GPT-6 Astra answers on the public API.
We sweep the public catalogs (OpenRouter, models.dev) and OpenAI’s own news feed every 60 seconds. The moment GPT-6 returns a real answer, your channels fire — email + Chrome push free, SMS and a phone call on paid tiers. No rumors, no reposted leaks: the alert means it’s answering.