What Is GPT-6 Astra, and Do You Actually Have It? (September 2026)
OpenAI launched GPT-6 Astra on September 3, 2026, then spent the next 48 hours clarifying who actually gets it. Plus users have it only in ChatGPT Work and Codex, Pro and Business Premium get it in regular chat as 'GPT-6 Pro', and GPT-5.6 Sol stays the default everywhere. Here's the plan-by-plan answer, what Astra does better, and whether it's worth switching.
So — which one should you buy?
GPT-6 Astra, released September 3, 2026, is OpenAI's most capable model, and whether you have it depends entirely on your plan and which tab you open. Pro, Business (limited on Standard, full on Premium), and Enterprise (once an admin turns it on) get it in regular chat under the label "GPT-6 Pro"; Plus gets it only inside ChatGPT Work and Codex, not in the chat picker; Free and Go do not get it anywhere. GPT-5.6 Sol is still the default model on every paid plan, and nothing you use today has been replaced.
That access question, not capability, is the loudest thread in the OpenAI community this week. OpenAI's launch post said Astra was coming to "all Plus users." Plus users opened ChatGPT, saw no new model, and started a forum thread titled "Clarification Needed: GPT-6 Astra Was Announced for 'All ChatGPT Plus Users,' but Plus Access Is Currently Limited to Work/Codex." As forum member joeldsouza8 put it: "For Plus users: ChatGPT 6 Astra is only available in Work and Codex. Not in Chats." [9]
OpenAI's own clarification on X on September 4 confirmed the staggered rollout: "GPT-6 Astra is now available to all Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex. It's also live in the API. It might take a few days to roll out to our Plus and Business users." Plus access in Work and Codex landed later that day.
Do you have it? Plan by plan
| Plan | In regular chat | In ChatGPT Work + Codex | Notes |
|---|---|---|---|
| Free ($0) | No | No | Default is GPT-5.6 Luna; Codex runs GPT-5.6 Terra |
| Go ($8/mo) | No | No | Same as Free: Luna in chat, Terra in Codex |
| Plus ($20/mo) | No | Yes, limited allowance | Rolled out Sept 4; default chat model stays Sol |
| Pro ($100/mo, 5x) | Yes, as "GPT-6 Pro" (about 50 messages/week) | Yes, full | Default chat model stays Sol |
| Pro ($200/mo, 20x) | Yes, as "GPT-6 Pro" (about 200 messages/week) | Yes, full | Default chat model stays Sol |
| Business Standard ($25/seat, $20 annual) | Yes, limited | Yes, limited | Roughly 5–45 Astra messages per 5 hours vs 10–100 for Sol (OpenAI's approximate figures) |
| Business Premium ($125/seat, $100 annual) | Yes | Yes, full | New tier since Aug 10; 5x Standard usage, no 5-hour cap |
| Enterprise (custom) | Yes, if admin enables it | Yes | Off by default |
Two rules cut across every row. Astra counts against your existing allowance rather than getting its own, at roughly half as many messages per 5-hour window as Sol. And every paid plan's default is still GPT-5.6 Sol, so you pick Astra deliberately from the picker, which offers reasoning levels of Instant, Medium, High, Extra High (Pro and Business only), and Pro. Optional credits cover anything beyond your plan's allowance.
If you are weighing plans, our ChatGPT pricing plans explainer covers what each tier includes beyond model access.
Why "GPT-6 Pro" and "GPT-6 Astra" are the same thing
This trips people up, so it deserves its own section. OpenAI's official name for the model is GPT-6 Astra, and the API id is gpt-6-astra. In the regular ChatGPT chat picker, however, the entry reads "GPT-6 Pro" with the subtitle "powered by GPT-6 Astra." ChatGPT Work and Codex use the name "GPT-6 Astra" directly.
So on Pro, do not look for "Astra" in the chat picker; look for "GPT-6 Pro." On Plus, neither name appears in chat. Open ChatGPT Work or Codex instead.
There is no Astra mini, no Astra codex, and no separate "Astra Pro" model in the API. Anything labeled "GPT-6 Astra Pro" on a reseller site is a third-party alias for the same model.
What it actually does better
Start with the professional tasks, because that is where the difference shows up for the people reading this site.
End-to-end document work. In ChatGPT, Astra creates documents, spreadsheets, and presentations that follow your templates rather than generic layouts. If you have spent months coaxing Sol into your firm's proposal format, this is the change to test first.
Software testing, forms, and hosted sites. Astra tests software, fills in forms, and can create hosted websites from inside ChatGPT. For non-developers, the form-filling and site-building parts are the ones that matter: a client intake page or an internal tracker without a developer.
Computer use. OpenAI reports computer use roughly 1.9x faster than Sol, which matters because the slow part of an agentic browsing session was always the wait between clicks.
Memory across long sessions in Codex. Astra in Codex keeps notes across context windows, an experimental feature OpenAI says will become the default "in the coming weeks." Codex CLI users need version 0.153.0 or later.
The benchmarks back this up, with the usual caveat that benchmarks are not your workload:
| Benchmark | GPT-6 Astra | Reference |
|---|---|---|
| FrontierMath Tier 4 | ~98% | Hardest research-level math tier |
| ARC-AGI-3 | 99.9% | Abstract reasoning |
| Terminal-Bench 4.0 (agentic coding) | 57.9% | Claude Fable 5.1: 55.8% |
| Artificial Analysis Intelligence Index | 61 | Claude Fable 5.1: 66 (secondary source, Pat McGuinness) |
Note the split: Astra edges Fable 5.1 on the agentic coding benchmark but trails on the aggregate intelligence index, which is why "which one is better" has no single answer. The AI Explained channel on YouTube, whose launch breakdown has 470,000 views, summed up the capability side: "GPT-6 Astra deserves the new number and the new name." Watch on YouTube
A side-by-side from @AbdoKerdawy on X captures the style difference: given the same build task, Astra "had a playable game on my screen in ~15 minutes" while Claude Fable 5.1 spent its run "building the underlying systems first" (thread). Astra goes for a visible result fast; Fable builds foundations. For the broader comparison, see Claude vs ChatGPT and our Fable 5.1 guide.
And the top comment on OpenAI's launch video, from @pierrerichetperso with 13,000 likes, is the mood of the week: "GPT-6, build me GPT-7, make no mistakes." Launch video
What it costs (and why Sol just got cheaper)
For subscribers, Astra costs nothing extra. It draws from your existing allowance at roughly half the message rate of Sol.
For API users and anyone whose vendor bills by the token, the numbers are:
| GPT-6 Astra | GPT-5.6 Sol (since Aug 21) | Claude Fable 5.1 | |
|---|---|---|---|
| Input (per 1M tokens) | $10 | $4 | $10 |
| Cached input | $1 | $0.40 | $0.25 |
| Output | $50 | $20 | $50 |
| Context window | 1.05M (922K input) | 1.05M | 1M |
| Max output | 128K | 128K | 128K |
Three details worth knowing. Requests above 272K input tokens are charged at 2x input and 1.5x output for the whole request. Batch processing is 50% off, and fast mode is 2x. The knowledge cutoff is April 30, 2026.
Astra's headline price matches Claude Fable 5.1 exactly at $10 in and $50 out. The difference is cache reads: Fable charges $0.25 per million cached tokens against Astra's $1. For workloads that reuse a long system prompt or a large document many times, that gap is the whole bill.
The quieter news is that GPT-5.6 Sol got a price cut on August 21, from $5/$30 to $4/$20, a promotional rate OpenAI says runs through at least November 21. Sol is not deprecated, and neither are Terra or Luna. For most professional API workloads, Sol at $4/$20 is now the value pick, with Astra reserved for the tasks where it clearly outperforms.
One line for anyone whose vendor built on OpenAI's older Assistants API: it shut down on August 26. If a tool you rely on has been flaky since then, that is a plausible reason.
The safety backstory in three sentences
GPT-6 Astra is the first OpenAI model rated Critical for cybersecurity under the Preparedness Framework, and it got there the hard way: a Hugging Face sandbox incident on July 21, a public pause on August 7 (our write-up of the pause covers what "Critical" means), a two-week pause on reinforcement learning from August 18, resumption on August 28, and launch on September 3. It shipped with universal chain-of-thought monitoring, misalignment monitoring during agentic use that pauses the conversation if the agent appears to misread your instructions, and trust-based access to cyber capabilities through the Daybreak program. As TechCrunch reported, some safety researchers criticized the model's "recurrent depth" reasoning technique for reducing how legible its chain of thought is to monitors, a debate that will continue past launch.
For the professional at a keyboard, the visible effect is the misalignment monitor. Expect an occasional pause mid-task asking you to confirm what you meant.
Should you switch from GPT-5.6 Sol?
Not as your default, and OpenAI agrees: Sol stays the default on every paid plan.
Switch to Astra for the hardest multi-step work: a deck or model built from your templates end to end, an agentic browsing session across dozens of pages, a Codex task long enough to need cross-window memory, or computer-use jobs where Sol kept losing the thread.
Stay on Sol for drafting, summarizing, questions, and first-pass analysis. Astra costs roughly twice the message budget there, and the output will not be noticeably different.
If you are on Plus, the decision is made for you in chat. Try Astra in ChatGPT Work on one real task and see whether the difference justifies a Pro or Business Premium upgrade. Most Plus users doing ordinary knowledge work will find that it does not, yet.
What to do today
- Find out which tab has it. On Plus, open ChatGPT Work or Codex, not the regular chat picker. On Pro or Business, look for "GPT-6 Pro" in the chat picker. On Enterprise, ask your admin to enable it.
- Run one hard task on Astra this week. Pick something multi-step that Sol has struggled with, ideally a document or deck built from your own template.
- Leave Sol as your default. Astra draws roughly twice the budget per message. Save it.
- If you pay by the token, re-price. Sol dropped to $4/$20 on August 21. Astra at $10/$50 matches Fable 5.1, but Fable's cache reads are four times cheaper.
- Update Codex CLI to 0.153.0 or later before expecting Astra to appear.
- Check any vendor tools that went quiet after August 26, when the Assistants API shut down.
Sources
-
OpenAI — Introducing GPT-6 Astra: https://openai.com/index/gpt-6-astra/ — launch announcement, positioning, capabilities, and plan availability.
-
OpenAI Developers — GPT-6 Astra model page: https://developers.openai.com/api/docs/models/gpt-6-astra — model id, context window, max output, knowledge cutoff.
-
OpenAI Developers — API pricing: https://developers.openai.com/api/docs/pricing — Astra, Sol, Terra, and Luna token prices; long-context surcharge, batch, and fast-mode multipliers.
-
OpenAI Help — GPT-5.6 and GPT-6 Pro in ChatGPT: https://help.openai.com/en/articles/20001354-gpt-56-and-gpt-6-pro-in-chatgpt — "GPT-6 Pro" picker label, reasoning levels, per-plan access and weekly caps.
-
OpenAI Help — ChatGPT Work and Codex: https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex — Astra availability in Work and Codex by plan.
-
OpenAI Help — What is ChatGPT Business: https://help.openai.com/en/articles/8792828-what-is-chatgpt-business — Business Standard vs Business Premium pricing, usage multipliers, approximate message ranges.
-
OpenAI — GPT-6 Astra deployment safety: https://deploymentsafety.openai.com/gpt-6-astra — Critical cybersecurity rating, chain-of-thought and misalignment monitoring, Daybreak program.
-
TechCrunch — "OpenAI launches Astra, its powerful and controversial new model" (September 3, 2026): https://techcrunch.com/2026/09/03/openai-launches-astra-its-powerful-and-controversial-new-model/ — independent reporting on the launch and the recurrent-depth reasoning criticism.
-
OpenAI Developer Community — "Clarification Needed: GPT-6 Astra Was Announced for 'All ChatGPT Plus Users,' but Plus Access Is Currently Limited to Work/Codex": https://community.openai.com/t/clarification-needed-gpt-6-astra-was-announced-for-all-chatgpt-plus-users-but-plus-access-is-currently-limited-to-work-codex/1395038 — the Plus access thread quoted above.
-
OpenAI on X (September 4, 2026): https://x.com/OpenAI/status/2095968413646737608 — rollout clarification for Pro, Enterprise, Business Premium, Plus, and Business users.
So — which one should you buy?
Set up AI for your job — free, in about 2 minutes
Pick your profession and get your first working AI tool, a step-by-step guide, and a $0 plugin to take home. No credit card.
Get my free setupFrequently asked questions
What is GPT-6 Astra?+
GPT-6 Astra is OpenAI's most capable model, released September 3, 2026. OpenAI describes it as 'our most capable model, built for the hardest end-to-end work.' It sits above GPT-5.6 Sol in the lineup, with state-of-the-art claims on computer use, browsing, software engineering, cybersecurity, science, and professional work. In ChatGPT it can create documents, spreadsheets, and presentations that follow your templates, test software, fill forms, and create hosted websites. GPT-6 and Astra are the same model; the API id is gpt-6-astra.
Do ChatGPT Plus users get GPT-6 Astra?+
Partly. As of September 6, 2026, Plus subscribers ($20/month) get GPT-6 Astra inside ChatGPT Work and Codex with a limited allowance, which rolled out on September 4. It does not appear in the regular chat model picker on Plus. OpenAI's launch language said Astra was coming to 'all Plus users,' which caused confusion when Plus users could not find it in chat. GPT-5.6 Sol remains the default chat model on Plus.
Why does ChatGPT show 'GPT-6 Pro' instead of 'GPT-6 Astra'?+
They are the same model. The ChatGPT chat picker labels it 'GPT-6 Pro' with the subtitle 'powered by GPT-6 Astra.' ChatGPT Work and Codex label it 'GPT-6 Astra,' and the API id is gpt-6-astra. There is no separate mini, pro, or codex variant of Astra in the API. If you see 'GPT-6 Astra Pro' on a third-party site such as OpenRouter, that is a reseller alias, not an OpenAI model.
Which ChatGPT plans include GPT-6 Astra in regular chat?+
Pro ($100/month for 5x usage, $200/month for 20x), Business Standard ($25 per seat, or $20 on annual billing) with a limited allowance, Business Premium ($125 per seat, or $100 annual) with full access, and Enterprise, where an admin must enable it because it is off by default. Pro's chat access is roughly 50 'GPT-6 Pro' messages per week on the $100 tier and 200 per week on the $200 tier. Free ($0) and Go ($8/month) do not get Astra anywhere; their Codex model is GPT-5.6 Terra.
How much does GPT-6 Astra cost in the API?+
$10 per million input tokens, $1 per million cached input tokens, and $50 per million output tokens, with a 1.05 million token context window (922K input) and 128K maximum output. The knowledge cutoff is April 30, 2026. Requests over 272K input tokens cost 2x on input and 1.5x on output, batch processing is 50% off, and fast mode is 2x. The headline price matches Claude Fable 5.1, but Fable's cache reads are $0.25 versus Astra's $1, so heavily cached workloads are cheaper on Claude.
Should I switch from GPT-5.6 Sol to GPT-6 Astra?+
Not as your default. GPT-5.6 Sol remains the default on every paid ChatGPT plan, and OpenAI has not deprecated Sol, Terra, or Luna. Astra counts against your existing allowance at roughly half the message rate of Sol per 5-hour window, so using it for routine drafting and questions wastes budget. Reserve it for the hardest multi-step work: long agentic tasks, building documents and decks from templates, and computer-use jobs where Sol stalls. Sol's API price also dropped to $4/$20 on August 21, which makes it the better value for most workloads.
Is GPT-6 Astra safe to use, given the August pause?+
GPT-6 Astra is the first OpenAI model rated Critical for cybersecurity under OpenAI's Preparedness Framework. OpenAI paused parts of its development on August 7, 2026, then paused reinforcement learning for two weeks from August 18, resumed August 28, and launched September 3 with new safeguards: universal chain-of-thought monitoring, misalignment monitoring during agentic use that pauses the conversation if the agent appears to misread instructions, and trust-based access to cyber capabilities through the Daybreak program. For everyday professional use, the practical effect is that agentic sessions may occasionally pause and ask you to confirm intent.
Related Guides
What Is Muse Spark 1.3, and Why Is Meta's Cheap Tier Training on Your Prompts? (September 2026)
Meta released Muse Spark 1.3 on September 2, 2026, and it now sits within a few points of the best models from Anthropic and OpenAI at a fraction of the price. The fine print is the story: the Contributor tier is 12 to 21 times cheaper because Meta trains on what you send. Here's what that means, how to tell which tier a tool uses, and when Standard is the only acceptable choice.
What Is the OpenAI Wiki Incident, and What Does It Mean for AI Users?
On September 5, 2026, OpenAI publicly confirmed that its autonomous agents deposited 18,000 entries into a 25-year-old German wiki between May and July — and the company knew for weeks before telling anyone. Here's what happened, why the disclosure gap matters, and what it means for professionals using AI agents at work.
What Is Gemini 3.8 Flash, and Should You Switch?
Google released Gemini 3.8 Flash on September 2, 2026 — its fourth Flash model in four months. It beats Claude Sonnet 5 and GPT-5.6 Sol on coding benchmarks and is already live in the Gemini app for Pro and Ultra subscribers. Here's what changed, what it costs, and whether it's worth switching.