OpenAI posted GPT-6 Astra's public rollout at 3:32pm ET; API docs put the model ID, $10/$50 pricing, and Sep 3 changelog on the same day as the X thread.

GPT-6 Astra rolls out: computer-use flagship, $10/$50, Trusted Access first

Generated by Grok Bot research on 2026-09-03. Native X + fetch ladder (developers.openai.com markdown). Treat as raw material — review before promoting.

Summary

At 3:32pm ET on 3 Sep 2026, @OpenAI posted that GPT-6 Astra is rolling out: limited organizations today, then ChatGPT Plus, Pro, Business, and Enterprise plus the OpenAI API and AWS over the coming days. The same-day API changelog and model card give the durable product facts: model ID gpt-6-astra, 1,050,000 context, $10 input / $50 output per 1M tokens at standard short context, and no none reasoning effort. The X thread's benchmark list (FrontierMath Tier 4, ARC-AGI 3, TerminalBench-4.0, Agents' Last Exam, AutomationBench, ScreenSpot Pro) is issuer-claim only — this pass did not retrieve the marketing page's score table. The thread opener is video and was not transcribed.

Findings

Availability: Trusted Access today, paid ChatGPT and API "in the coming days"

The load-bearing availability sentence is in the official thread, with the truncated line completed by note_tweet: GPT-6 Astra "is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS." The same post tells people to get the ChatGPT desktop app.

The API model card and Using GPT-6 Astra repeat the rollout in developer language: "rolling out today for enterprises in our Trusted Access Program, with access through API and our Plus, Pro, Business and Enterprise plans coming in the coming days." That is slightly narrower than the X post (X names AWS; the model card names Trusted Access and does not mention AWS). Treat AWS as an X-only claim until an issuer docs page states it.

The Trusted Access href on those docs pages currently points at https://openai-dotcom-preview.vercel.app/form/enterprise-trusted-access-for-cyber/ — a Vercel preview host, not openai.com. Record it as retrieved, not as a durable form URL.

The Sep 3 changelog is dated, not just "today": "Released GPT-6 Astra, our most capable model, built for the hardest end-to-end work," on Chat Completions and Responses.

This is a different event from the 1–2 Sep Preparedness writeups already in the thread (path-to-astra / Critical cyber). Today's docs still point at Path to Astra for misalignment monitoring. Do not collapse "available soon / testers" into "GA."

Product shape: computer-use flagship, not a cheap Flash-class model

@OpenAI's second post calls GPT-6 Astra "the most intelligent and aligned model in the world" and points at https://openai.com/index/gpt-6-astra/. Isolated WebFetch timed out on that URL; curl returned a JS/cookie challenge (403 then 500). This pass did not retrieve the marketing article. Cite it as a named pointer. Hydration is the developers.openai.com markdown.

What did retrieve:

Model ID gpt-6-astra. Input: text and image. Output: text. 1,050,000 context window, 922,000 max input tokens, 128,000 max output tokens, knowledge cutoff Apr 30 2026.
reasoning.effort supports low, medium, high, xhigh, and max. No none. No custom temperature / top_p / logprobs. Tool calling requires the Responses API.
New API surface on the same day: async tool calling, mid-turn steering over WebSockets, and mid-conversation configuration_update for reasoning effort while keeping the cached prefix.
Misalignment monitoring on supported Responses requests, which can alert or stop a conversation. Fast mode is unavailable for GPT-6 Astra with EU data residency.
Supported Responses tools on the model card include computer_use, hosted_shell, apply_patch, skills, mcp, and tool_search.

The model guide's cost claim is qualitative and cites the unrecovered marketing page: Astra "achieves stronger results while using substantially fewer output tokens—delivering a lower estimated API cost per task than earlier models despite its higher per-token pricing." Hold that as an OpenAI estimate, not a measured dollar figure.

Pricing: $10 / $50, same list band as Claude Fable 5.1, 2.5× Sol

The model card and pricing table agree on standard short-context text tokens:

Tier	Input	Cached input	Cache writes	Output
Standard short	$10.00	$1.00	$12.50	$50.00
Standard long (>272K input; 2× in / 1.5× out for the full request)	$20.00	$2.00	$25.00	$75.00
Batch / Flex short	$5.00	$0.50	$6.25	$25.00
Fast mode short	$20.00	$2.00	$25.00	$100.00

Units are per 1M tokens. Regional processing (data residency) adds 10% for models released on or after 5 Mar 2026.

On the same pricing page, gpt-5.6-sol is $4.00 / $20.00 standard short. Astra is 2.5× that list price. This morning's already-filed X clip put Claude Fable 5.1 at $10 / $50 — Astra's standard short-context list matches that band, it does not undercut it. Qwen3.8-Max-0902's already-filed $2 / $6 remains a different price class.

Benchmark names, not scores

The X thread names boards and does not print scores in the post text (charts are images; this pass did not OCR them):

Agents' Last Exam, AutomationBench, and ScreenSpot Pro as "computer workflow tasks across professions."
FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0, plus Terminal-Bench Science 0.1 and HealthBench Pro for "scientific discovery."
"Most aligned model," with "substantial improvements in understanding user intent" — no number.

Do not transcribe chart pixels. Independent board snapshots are not in this pass.

Contradictions and open questions

Marketing page not retrieved. https://openai.com/index/gpt-6-astra/ is the URL OpenAI attached to the thread. Isolated fetch timed out; curl hit a JS/cookie wall. Scores, system card, and any "lower cost per task" table live there until a later fetch.
X vs docs on who gets it today. X: limited organizations today, paid ChatGPT + API + AWS "over the coming days." Docs: Trusted Access Program today, API and Plus/Pro/Business/Enterprise "in the coming days." AWS is X-only in this pass.
Trusted Access href is a Vercel preview. Retrieved as https://openai-dotcom-preview.vercel.app/form/enterprise-trusted-access-for-cyber/. Do not treat that host as the long-term form.
"Most intelligent / most aligned in the world" is issuer rhetoric. Carry it as OpenAI's claim. The named boards are listed; the numbers are not.
Thread opener is video. https://x.com/OpenAI/status/2095595741528125780 ("Anything you can do on a computer, Astra can do for you. Fast.") was not transcribed. Do not promote that permalink until /transcribe-clipping.
Already in the thread, deliberately not re-filed: Gemini 3.8 Flash (morning), K2 Horizon / Qwen3.8-Max-0902 / Mostik (10am X pass), WeatherNext 3 (1pm). Anthropic, Karpathy, and GeminiApp had zero original posts after 10am ET. Grok news clusters on Martian routing, Baseten Base Labs, and ngrok.ai were not hydrated.

Provenance

Method: Grok Bot / native X (search_news discovery, get_users_posts / get_posts_by_ids for text) / fetch ladder (developers.openai.com .md via curl). Isolated WebFetch timed out on openai.com. search_posts_all skipped. Generated: 2026-09-03 Rounds: 1 of 3 — early-exit: one issuer rollout, docs hydrated, remaining clusters were recap or already filed. URLs fetched: 6 successful (X thread + developers markdown), 2 failed (openai.com index 403/timeout; guessed system-card PDF 404) Fetch notes: developers.openai.com/api/docs/models/gpt-6-astra.md, .../guides/latest-model.md?model=gpt-6-astra, .../pricing.md, .../changelog.md retrieved as markdown. openai.com/index/gpt-6-astra/ not retrieved. Playco/Legora customer stories appeared in WebSearch only — not fetched, not cited as grain.

Web sources:

GPT-6 Astra model card — ID, context, pricing, tools, Trusted Access wording.
Using GPT-6 Astra — alignment/computer-use framing, new API features, misalignment monitoring pointer.
OpenAI API changelog, Sep 3 2026 — dated release plus async tools / steering / reasoning-update.
OpenAI API pricing — standard / batch / flex / fast tables including gpt-6-astra vs gpt-5.6-sol.
GPT-6 Astra marketing page (named, not retrieved) — pointer from X and the model guide.

X sources:

X post by @OpenAI (2026-09-03) — "most intelligent and aligned," SOTA claim across computer use / browsing / SWE / cyber / science; links openai.com/index/gpt-6-astra/.
X post by @OpenAI (2026-09-03) — Agents' Last Exam, AutomationBench, ScreenSpot Pro.
X post by @OpenAI (2026-09-03) — "most aligned," user-intent claim.
X post by @OpenAI (2026-09-03) — FrontierMath Tier 4, ARC-AGI 3, TerminalBench-4.0, Terminal-Bench Science 0.1, HealthBench Pro.
X post by @OpenAI (2026-09-03) — rollout: limited orgs today; Plus/Pro/Business/Enterprise + API + AWS coming days; ChatGPT desktop.
Video, not cited as grain: X post by @OpenAI (2026-09-03) — thread opener with video. No transcript. fetch_method: x-mcp.
Negative: @AnthropicAI, @karpathy, @GeminiApp — zero original posts after 2026-09-03T14:00:00Z. @GoogleDeepMind only returned the already-filed WeatherNext 3 thread.

Grokipedia:

Not used.