Vintage: 2026-09. Primary evidence is OpenAI's Sep 3, 2026 API changelog + model card + pricing table (developers.openai.com markdown) plus official @OpenAI X text, as hydrated in 2026-09-03-gpt-6-astra-rolls-out-computer-use-flagship. The marketing page openai.com/index/gpt-6-astra/ was not retrieved (timeout / JS-cookie wall). The thread opener (2095595741528125780) is video and was not transcribed — do not treat that permalink as grain. Capability and price claims are a snapshot of that day's fetched docs, not a later card. 2026-09-04 overnight: staff banked-reset + messy-rollout apology + Azure Foundry Limited Access in 2026-09-04-x-overnight-opencode-omen-alpha-unsloth-glm-5-3-flash. 2026-09-07 weekend clip: Sep 4 issuer text in 2026-09-07-weekend-x-ai-meta-aira3-kaggle-gold-openai-wiki — Pro / Enterprise / Business Premium + API live; Plus/Business then out per @sama. Do not re-litigate $10/$50 or benches from the Sep 3 source. Do not collapse ChatGPT consumer access with Foundry enterprise Limited Access. 2026-09-07 overnight clip: staff Astra-low vs Sol-high calibration + issuer −59.2% GPU note in 2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi — not GA, not a score table. 2026-09-07 18:14 UTC AA clip: artificial-analysis-intelligence-index v4.3 Index 53 (max) + effort-labeled TB4 / AutomationBench prints in 2026-09-07-artificial-analysis-intelligence-index-v4-3 — AA, not an OpenAI card. Still not GA. Do not rewrite $10/$50. 2026-09-09 overnight clip: issuer “fully rolled out” in Codex + ChatGPT Work + gpt-tv + AA artificial-analysis-model-release-pages in 2026-09-09-overnight-x-openai-astra-full-work-codex-rollout-aa-model. Chat-picker vs Work/Codex left open. Do not invent GA. 2026-09-10 afternoon: chatgpt-for-financial-services cites Astra as the reasoning model in a tailored ChatGPT Work offering — not a SKU rewrite. Do not invent GA or re-litigate $10/$50.

GPT-6 Astra

One-line summary: OpenAI's September 3, 2026 computer-use flagship (gpt-6-astra) — Trusted Access today, paid ChatGPT and API "in the coming days," $10 / $50 standard short-context list. Not GA. A different event from the Sep 1 Path to Astra / Critical cyber writeup.

What it is

A dated product rollout, not the Sep 1–2 Preparedness writeup on openai-astra-critical-cyber. The Sep 3 changelog says "Released GPT-6 Astra, our most capable model, built for the hardest end-to-end work," on Chat Completions and Responses. The model card gives model ID gpt-6-astra, 1,050,000 context, and the $10 / $50 standard short-context list that the pricing table repeats.

This page records issuer-hydrated claims from fetched developers.openai.com markdown and official X text. It does not re-file same-day gemini-3-8-flash, k2-horizon, qwen-3-8-max-0902, mostik-ai, or weathernext-3. It does not cite the untranscribed thread-opener video.

Why it matters to this thread

Frontier model releases, pricing, and computer-use / agentic deployment are in-scope. This is the first dated GPT-6 Astra SKU the thread has as a citable product card. Distinct from openai-astra-critical-cyber (Sep 1 Critical cyber designation; "available soon" / testers / Daybreak Blue). Do not collapse Trusted Access / "coming days" into GA.

Key facts (from 2026-09-03-gpt-6-astra-rolls-out-computer-use-flagship)

Availability: Trusted Access today — not GA

From 2026-09-03-gpt-6-astra-rolls-out-computer-use-flagship (official @OpenAI X, 2026-09-03, 2095595757072191802, 3:32pm ET): 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." Same post tells people to get the ChatGPT desktop app.
From the same source (API model card + Using GPT-6 Astra): "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."
X vs docs, not flattened. X names AWS and "limited set of organizations." Docs name Trusted Access and do not mention AWS. Treat AWS as an X-only claim until an issuer docs page states it.
From the same source: 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. Retrieved as written; not a durable form URL.
From the same source (Sep 3 changelog): dated release on Chat Completions and Responses — "Released GPT-6 Astra, our most capable model, built for the hardest end-to-end work."
Today's docs still point at Path to Astra for misalignment monitoring. That is a pointer, not a rewrite of openai-astra-critical-cyber.

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

From the same source (@OpenAI, 2026-09-03): 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; curl hit a JS/cookie challenge (403 then 500). Marketing article not retrieved. Cite the URL as a named pointer. Hydration is the developers markdown.
From the same source (model card): 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.
From the same source (changelog): reasoning.effort supports low, medium, high, xhigh, and max. No none. No custom temperature / top_p / logprobs. Tool calling requires the Responses API.
From the same source (changelog): same-day API surface — async tool calling, mid-turn steering over WebSockets, and mid-conversation configuration_update for reasoning effort while keeping the cached prefix.
From the same source (model guide): misalignment monitoring on supported Responses requests can alert or stop a conversation. Fast mode is unavailable for GPT-6 Astra with EU data residency.
From the same source (model card): supported Responses tools include computer_use, hosted_shell, apply_patch, skills, mcp, and tool_search.
From the same source (model guide, qualitative, 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 as an OpenAI estimate, not a measured dollar figure.

Pricing: $10 / $50 standard short — matches Fable, 2.5× Sol

From the same source (model card + pricing table), standard short-context text tokens, per 1M 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

From the same source (same pricing page): regional processing (data residency) adds 10% for models released on or after 5 Mar 2026.
From the same source (same pricing page): gpt-5.6-sol is $4.00 / $20.00 standard short. Astra is 2.5× that list price.
From the same source (comparison to already-filed pages, not a re-file): this morning's X clip put claude-fable-5-1 at $10 / $50 — Astra's standard short-context list matches that band, it does not undercut it. qwen-3-8-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). Independent board snapshots are not in this pass.

From the same source (2095595744300503356): Agents' Last Exam, AutomationBench, and ScreenSpot Pro as "computer workflow tasks across professions."
From the same source (2095595752815030713): FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0, plus Terminal-Bench Science 0.1 and HealthBench Pro for "scientific discovery."
From the same source (2095595748528452037): "Most aligned model," with "substantial improvements in understanding user intent" — no number.

Do not transcribe chart pixels. Do not invent scores. See ai-coding-benchmarks.

Overnight follow-through: messy ChatGPT access + Azure Foundry Limited Access (2026-09-04)

From 2026-09-04-x-overnight-opencode-omen-alpha-unsloth-glm-5-3-flash — not a rewrite of the Sep 3 Trusted Access / coming-days docs. Not GA.

From the same source (OpenAI staff @thsottiaux, 2026-09-03 23:12 UTC): “one banked reset for every day you don't have access to Astra on your paid ChatGPT plan, starting today.”
From the same source (@sama, 2026-09-04 01:02 UTC): apologized for the “messy rollout”; “should be able to begin broad rollout to API customers and chatgpt subscribers in the near future,” starting with Pro.
From the same source (@satyanadella, 2026-09-04 03:21 UTC) linked the Azure blog; @sama quote-posted “We are also excited!”
From the same source (issuer Azure blog, dated September 3; fetched 2026-09-04; Schema.org datePublished 2026-09-03T18:15:00+00:00): Astra “begins rolling out today through the Microsoft Foundry Limited Access Program, with availability expanding to participating customers over the coming days.” Emphasizes multi-step planning, polished artifacts, and computer use across applications, with Foundry identity/governance controls; notes OpenAI reports SOTA on selected computer-use evals and that prompts/outputs are not used to train the models.
Official @OpenAI had zero original posts in the overnight window (start_time 2026-09-03T22:00:00Z).

Channels differ. Staff banked-reset + “near future” broad ChatGPT/API rollout is consumer / API. Azure “Limited Access Program… coming days” is Foundry enterprise. Do not collapse them. Full Microsoft attach: microsoft.

Weekend follow-through: Pro / API live; Plus/Business then out (2026-09-04)

From 2026-09-07-weekend-x-ai-meta-aira3-kaggle-gold-openai-wiki — incremental availability only. Do not re-litigate pricing or benchmarks.

From the same source (official @OpenAI, 2026-09-04): GPT-6 Astra “is available to all Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex, and live in the API,” with Plus/Business rollout still pending.
From the same source (@sama): mirrored that availability note.
From the same source (@sama): then said Plus/Business were out.

This updates the Sep 3 Trusted-Access / “coming days” filing and the overnight “near future, starting with Pro” note. It does not rewrite the fetched $10/$50 card or named boards. Pachocki’s 6 Sep essay claim that Astra is “significantly better aligned than GPT-5.6 Sol” is filed on openai-alien-mind — issuer essay, not a new card.

Staff calibration + issuer GPU cut (2026-09-06/07) — not a GA rewrite

From 2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi — primary X + issuer, not a product card. Still not GA.

From the same source (OpenAI staff @thsottiaux, 2026-09-06): “GPT-6 Astra on low performs better than GPT-5.6 Sol on high”; users happy on Sol high should move to low or medium on Astra. Staff calibration guidance, not an official bench table. Adjacent: gpt-5-6-sol-pricing.
From the same source (same account, 2026-09-05): while Astra was not generally available it was “probably our biggest competitive advantage”; internal productivity jumped enough that some plans shifted 6 months earlier to ship at DevDay instead of mid next year. Staff claim; keep attributed.
From the same source (2026-09-06 follow-up): usage improvements for power users logged in with a ChatGPT account; “up to 3–4X less usage” drawn from subscription on the long tail; “no change in quality.” Staff claim. Secondary @AndrewCurran_ is commentary, not the OpenAI blog.
From the same source (@JensenHuang, 2026-09-06): “GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.” Nvidia CEO rhetoric, not a lab capability paper. Keep the AGI label contested — agi-definitions-and-benchmark-saturation.
From the same source (issuer research-acceleration page): August 7 preliminary Astra critical-cyber flag → following week Astra-class GPU allocation −59.2%, other classes +17.2%. Full treatment: openai-research-acceleration / openai-astra-critical-cyber. Not a SKU rewrite.

AA Intelligence Index v4.3 (2026-09-07 18:14 UTC) — not an OpenAI card

From 2026-09-07-artificial-analysis-intelligence-index-v4-3 — independent AA thread + fetched pages. Not GA. Not a rewrite of $10/$50. Sep 3 named boards now have AA prints; keep effort labels.

From the same source (Astra model page + announce): Index v4.3 53 (max), tied with claude-fable-5-1 (max with fallback). Cost-per-Index-task $3.26 (model page repeats $3.26; cost post vs Fable $7.63, 57% lower).
From the same source (TB4 thread): Terminal-Bench v4.0 Astra (max) 59.1%; page blurb also lists Astra (xhigh) 59.6%. Same max figure. 19 percentage points ahead of GPT-5.6 Sol (max) 39.9%.
From the same source (AutomationBench thread + page): Astra (max) Score 68.5%; page also lists Astra (xhigh) 67.2%. Full-complete / no-violation 41.6% of workflows. Full treatment: automationbench-aa, artificial-analysis-intelligence-index.

Overnight: “fully rolled out” in Codex + ChatGPT Work (2026-09-08/09)

From 2026-09-09-overnight-x-openai-astra-full-work-codex-rollout-aa-model — issuer X text. Do not invent GA. Do not flatten chat-picker vs Work/Codex.

From the same source (@OpenAI, 2026-09-08 21:06:10Z): “Astra is fully rolled out to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work.” Points to live inspiration at openai.com/gpt-tv/. Full treatment of the demo surface: gpt-tv.
From the same source (issuer GPT-TV; clip curl 200 after WebFetch 403): title GPT-TV | OpenAI; marketing copy includes “Powered by GPT-6 Astra”; livestream/player chrome. Product surface, not a model card.
Secondary NotebookCheck says Plus Astra may remain Work/Codex-only while chat “GPT-6 Pro” stays higher tiers. Not grain for plan gates. OpenAI’s overnight post does not mention the chat picker. Leave the surface split open.

AA Model Release pages (2026-09-09) — still not an OpenAI card

From the same source — independent AA product/UI + fetched pages. Not a rewrite of $10/$50 or of Index v4.3 max 53. Keep effort-range lines split. Full treatment: artificial-analysis-model-release-pages.

From the same source (@ArtificialAnlys, 2026-09-09 03:26:01Z): Model Release pages compare intelligence, cost, and speed across effort levels. Worked example (AA, not OpenAI): Astra 46–53 Intelligence Index and 1.6–8.2 minutes per task vs claude-fable-5-1 47–53 on 4.2–12.2 minutes.
From the same source (releases index + 2097526979427971324): Astra listed OpenAI · Sep 2026 · 6 variants; Intelligence 45–53; cost/task $0.82–$3.26; 1M context. 45–53 vs 46–53 not flattened.
From the same source (GPT-6 Astra (max)): AA 53 Intelligence Index at max; $3.26/task; $10/$50 per 1M in/out; 1M context; knowledge cutoff Apr 30 2026; released Sep 3 2026. Max 53 / $3.26 matches the already-filed v4.3 print; $10/$50 is AA repeating a list already on this page — do not re-litigate.

Named as the reasoning model in ChatGPT for Financial Services (2026-09-10 afternoon)

From 2026-09-10-openai-chatgpt-for-financial-services-pairs-astra — issuer product page + official X text. Not a SKU rewrite. Do not invent GA. Do not re-litigate $10/$50. Chat-picker vs Work/Codex stays open.

From the same source (Introducing ChatGPT for Financial Services, dated September 10, 2026 + @OpenAI): the FS offering “combines built-in financial data with GPT‑6 Astra’s reasoning.” Full treatment: chatgpt-for-financial-services. Distinct from midday chatgpt-work-data-agent. Announce video transcript is . — not grain.

What this source does not establish

Availability is dated issuer text, not a re-measured GA. Sep 3: Trusted Access / coming days. Overnight: banked resets + “near future” (starting Pro) + Foundry Limited Access. Weekend clip (Sep 4): Pro / Enterprise / Business Premium + API live; Plus/Business then out per @sama. 2026-09-09 overnight: “fully rolled out” in Codex + ChatGPT Work — chat-picker still open. Do not collapse ChatGPT consumer access with Foundry enterprise Limited Access. Do not invent a “GA” date.
Not a rewrite of openai-astra-critical-cyber. Sep 1 Path to Astra / Critical cyber / ExploitBench split / Daybreak Blue stay on that page. Today's docs still point at that post for misalignment monitoring.
Marketing page not retrieved. Scores, system card, and any "lower cost per task" table live at openai.com/index/gpt-6-astra/ until a later fetch.
AWS is X-only in this pass.
Trusted Access href is a Vercel preview, not a durable openai.com form.
"Most intelligent / most aligned in the world" is issuer rhetoric. Named boards are listed; numbers are not.
Thread opener is video (2095595741528125780). No transcript. Do not cite as grain.
Cost-per-task is an OpenAI estimate citing the unrecovered marketing page — not a measured dollar figure.
Not a re-file of gemini-3-8-flash, k2-horizon, qwen-3-8-max-0902, mostik-ai, or weathernext-3.
Playco / Legora customer stories appeared in WebSearch only — not fetched, not cited as grain.
No ticker, 8-K, or stock-market tag.
Sep 6–7 staff/issuer attach is not GA and not a score table. Tibo calibration + usage drawdown are attributed staff posts. Jensen “AGI has arrived” is definitional rhetoric. Issuer −59.2% is a one-week GPU-allocation note, not a product-card change. From 2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi.
AA Index 53 / $3.26 is Artificial Analysis, not an OpenAI card. Do not back-fill those figures onto the Sep 3 marketing page. From 2026-09-07-artificial-analysis-intelligence-index-v4-3.
Sep 8 Buckmaster statement names Astra “only used for writeups and auditing” on the forced-blowup program — not a SKU / access / bench rewrite. From 2026-09-08-x-morning-buckmaster-alpoge-ai-fluid-proofs-openai-credit. See buckmaster-alpoge-forced-blowups.
Sep 8 afternoon OpenAI NS claim names Astra for ~17h Lean formalization of the forced-NS writeup; the NS model itself is an unnamed internal model claimed “significantly more capable than GPT‑6 Astra.” Not a SKU / access / bench rewrite; not a named next SKU. From 2026-09-08-x-afternoon-openai-navier-stokes-images-2-5. See openai-forced-navier-stokes-claim.
Overnight “fully rolled out … in Codex and ChatGPT Work” is issuer X wording, not a measured chat-picker GA. Secondary NotebookCheck says Plus may stay Work/Codex-only. From 2026-09-09-overnight-x-openai-astra-full-work-codex-rollout-aa-model.
Afternoon chatgpt-for-financial-services is a Work product cite, not a model-card or GA event. Astra is named as the reasoning model. Do not invent prices or close the chat-picker question. From 2026-09-10-openai-chatgpt-for-financial-services-pairs-astra.
SMB “16 plugins” video is flagged, not grain. Do not invent plugin names. Same overnight clip; not a SKU rewrite.
AA Model Release effort ranges are Artificial Analysis, not an OpenAI card. Worked-example 46–53 vs listing 45–53 kept split. From 2026-09-09-overnight-x-openai-astra-full-work-codex-rollout-aa-model.

Contradictions / tensions

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. Not reconciled.
Overnight ChatGPT vs Foundry channels. Staff banked-reset + @sama “near future” broad API/ChatGPT rollout vs Azure “Limited Access Program… coming days.” Different surfaces. Not collapsed. From 2026-09-04-x-overnight-opencode-omen-alpha-unsloth-glm-5-3-flash.
Staff “Astra low > Sol high” vs $10/$50 list. Calibration guidance from @thsottiaux, not a fetched bench. From 2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi. Not collapsed into the pricing table.
Trusted Access href is a Vercel preview (openai-dotcom-preview.vercel.app). Retrieved, not treated as the long-term form.
Sep 1 "available soon / testers" vs Sep 3 Trusted Access / coming days. Chronological refinement of access posture — still not GA. Do not flatten the two events into one launch.
"Most intelligent / most aligned" vs no retrieved scores. Issuer claim; marketing score table not fetched.
Qualitative "lower cost per task" vs $10 / $50 list. Estimate cites the unrecovered marketing page; list price is from the fetched card + pricing table.
AA $3.26/Index-task vs $10/$50 list. Different units (composite Index task vs per-1M tokens). Not collapsed. From 2026-09-07-artificial-analysis-intelligence-index-v4-3.
Overnight “fully rolled out” vs Plus chat-picker. Issuer names Codex + ChatGPT Work; does not mention the chat picker. Secondary NotebookCheck says Work/Codex-only for Plus. Not flattened. From 2026-09-09-overnight-x-openai-astra-full-work-codex-rollout-aa-model.
AA worked-example Astra Intelligence 46–53 vs releases listing 45–53. Same overnight clip, two AA surfaces. Not reconciled.

Open questions

When do "coming days" become dated Plus / Pro / Business / Enterprise / API availability — and does AWS ever appear on an issuer docs page? Partially dated, not closed. From 2026-09-07-weekend-x-ai-meta-aira3-kaggle-gold-openai-wiki: Sep 4 @OpenAI said Pro / Enterprise / Business Premium + API were live, Plus/Business pending; @sama then said Plus/Business were out. 2026-09-09 overnight: @OpenAI says fully rolled out to Plus/Pro/Business/Enterprise in Codex and ChatGPT Work. Does that close the Plus chat-picker gap? Secondary NotebookCheck says Work/Codex-only for Plus; the overnight post does not mention the picker. Leave open. Foundry Limited Access and AWS-on-docs stay unresolved. Do not invent GA.
What does the unfetched marketing page add (scores, system card, cost-per-task table) that the developers markdown omits?
Does the Vercel-preview Trusted Access href get replaced by an openai.com form?
How, if at all, Trusted Access today relates to Daybreak Blue / testers on openai-astra-critical-cyber is not answered here. The existing "available soon → SKU" question on that page stays open.

Sources

2026-09-03-gpt-6-astra-rolls-out-computer-use-flagship
2026-09-04-x-overnight-opencode-omen-alpha-unsloth-glm-5-3-flash
2026-09-07-weekend-x-ai-meta-aira3-kaggle-gold-openai-wiki
2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi
2026-09-07-artificial-analysis-intelligence-index-v4-3 — AA Index v4.3 53 (max); TB4 / AutomationBench effort-labeled; $3.26/task. Not an OpenAI card.
2026-09-08-x-morning-buckmaster-alpoge-ai-fluid-proofs-openai-credit — Astra “only used for writeups and auditing”; not a SKU rewrite
2026-09-08-x-afternoon-openai-navier-stokes-images-2-5 — Astra ~17h Lean on OpenAI NS; internal model ≫ Astra; not a SKU rewrite
2026-09-09-overnight-x-openai-astra-full-work-codex-rollout-aa-model — issuer “fully rolled out” in Codex + ChatGPT Work; gpt-tv; AA Model Release ranges (45–53 vs 46–53 kept split). Chat-picker left open. SMB plugins video flagged, not grain.
2026-09-10-openai-chatgpt-for-financial-services-pairs-astra — issuer FS Work offering cites Astra as the reasoning model; not a SKU rewrite; video transcript not grain.

GPT-6 Astra

GPT-6 Astra

What it is

Why it matters to this thread

Key facts (from 2026-09-03-gpt-6-astra-rolls-out-computer-use-flagship)

Availability: Trusted Access today — not GA

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

Pricing: $10 / $50 standard short — matches Fable, 2.5× Sol

Benchmark names, not scores

Overnight follow-through: messy ChatGPT access + Azure Foundry Limited Access (2026-09-04)

Weekend follow-through: Pro / API live; Plus/Business then out (2026-09-04)

Staff calibration + issuer GPU cut (2026-09-06/07) — not a GA rewrite

AA Intelligence Index v4.3 (2026-09-07 18:14 UTC) — not an OpenAI card

Overnight: “fully rolled out” in Codex + ChatGPT Work (2026-09-08/09)

AA Model Release pages (2026-09-09) — still not an OpenAI card

Named as the reasoning model in ChatGPT for Financial Services (2026-09-10 afternoon)

What this source does not establish

Contradictions / tensions

Open questions

Sources

Related