GPT-6 Astra is OpenAI’s new flagship AI model, announced on 3 September 2026. It can operate a computer on your behalf, hold about 1.05 million tokens in a single conversation, and stay on a task that runs for hours instead of minutes.
It is also 2.5 times more expensive per token than the model it replaces, and independent testers found it got worse at a few jobs while getting much better at others.
This page walks through all of it in plain English, from the release date and pricing to the benchmarks, the computer use feature and the things Astra still gets wrong.
What is GPT-6 Astra?
GPT-6 Astra is the sixth-generation flagship model from OpenAI, released on 3 September 2026. It accepts text and images and replies in text only, with no audio or video output. Its headline change is that it can drive a computer directly, clicking, typing and filling in forms, rather than only writing answers back to you.
That shift matters more than the version number suggests. Earlier models were assistants you talked to. Astra is built to be handed a job. In OpenAI’s announcement, the company describes it filling in forms, updating CRM records, managing calendars, researching online, analysing data and building websites.
It also produces work in the shape people need it: documents, presentations, spreadsheets and analyses that follow a template you supply. OpenAI says the model has assisted with mathematics and health research, including open problems in prime gap theory. If you are placing Astra against the wider set of generative AI tools already in everyday use, that template-following output is the part that changes day-to-day work first.
The technical specification sits in the GPT-6 Astra model reference. Context window 1,050,000 tokens. Maximum output 128,000 tokens. Knowledge cutoff 30 April 2026, so it has no built-in awareness of anything after that date unless you give it search access or paste the information in.
Why did OpenAI name it Astra instead of just GPT-6?
OpenAI released the model under the name GPT-6 Astra, pairing the generation number with a codename in the same style as its predecessor, GPT-5.6 Sol. OpenAI has not published a detailed explanation of why it chose the name Astra, so anyone telling you what it stands for is guessing.
What can be said is what the naming pattern does. Keeping “GPT-6” preserves the version ladder people already understand, while “Astra” gives this release a handle that survives point upgrades. When GPT-6.1 arrives, “Astra” is what distinguishes this particular model in the picker.
For practical purposes, GPT-6, GPT6 and Astra all refer to the same thing right now. The API identifier is gpt-6-astra.
When was GPT-6 Astra released, and when do you get access?
GPT-6 Astra was announced on 3 September 2026. Access started with a limited set of organisations through a programme OpenAI calls Daybreak Access, and is extending to all ChatGPT Plus, Pro, Business and Enterprise users “over the coming days”. It is also available through the OpenAI API, Microsoft Azure and AWS Bedrock.
The staggered rollout has caused real confusion. Plus subscribers have been comparing notes about who can see the model and who cannot, because paying for the same plan does not currently guarantee the same model list. A missing model picker entry is expected during a phased release, not a fault with your account.
Business and Enterprise accounts have an extra step. Astra is off by default, and an administrator has to turn it on for the workspace. On a company plan, ask your IT or workspace admin rather than OpenAI support.
One thing OpenAI has not confirmed, as of 5 September 2026, is whether Free-tier ChatGPT users will get Astra at all. No date, no commitment either way. Our step-by-step guide to how to access ChatGPT Astra covers the plan-by-plan situation and what to check when the model does not show up.
What is genuinely new in GPT-6 Astra?
Four things stand out: computer use, long-horizon work that survives across sessions, noticeably stronger coding, and a context window of 1.05 million tokens. Everything else in the release is refinement. These four change what you can reasonably ask the model to do without supervising every step.
What does “computer use” actually mean?
Computer use means the model controls a computer interface directly rather than describing what you should click. OpenAI reports it completes tasks 47% faster per task than the previous generation. The listed examples are ordinary office work: form filling, CRM updates, calendar management, online research, data analysis, website building, and installing or troubleshooting software.
This is also the riskiest capability, because a model that can click things can click the wrong things. We cover the failure modes, the permission model and sensible guardrails in our breakdown of OpenAI Astra computer use and its risks.
How much difference does a 1.05 million token context window make?
Roughly, 1,050,000 tokens is around three-quarters of a million words held in view at once. A long codebase, a year of meeting notes or a full set of contracts fits in one conversation, without chopping it into pieces and hoping the model remembers the earlier chunks.
What does “long-horizon work” mean in practice?
It means tasks that run across sessions rather than finishing in one reply. Astra preserves context between sessions, keeps searchable notes during long runs, can ask questions asynchronously in Codex, and sticks more closely to stated task boundaries and environment restrictions. Independent testing on multi-week knowledge work showed the largest single gain of the release.
Is GPT-6 Astra better at coding?
OpenAI describes production-quality code needing fewer iterations, with better communication while it works. Independent testing supports the direction but not a lead over the field: Artificial Analysis scores it level with Claude Opus 5 and Fable 5 on coding agents, and 70% more token-efficient than GPT-5.6 Sol.
How does GPT-6 Astra score on benchmarks?
OpenAI’s own numbers are close to saturation on several tests, while independent testing from Artificial Analysis is more mixed. Read both, because they measure different things: OpenAI reports capability ceilings, Artificial Analysis reports how the model behaves across a broad basket of tasks at a given cost.
| Benchmark | Result | Source |
|---|---|---|
| ARC-AGI-3 | 99.9%, with human parity on 96% of levels | OpenAI |
| FrontierMath Tier 4 | 98% | OpenAI |
| ExploitBench | 100% (GPT-5.6 Sol scored 78.5%) | OpenAI |
| Computer use speed | 47% faster per task than the previous generation | OpenAI |
| Alignment testing | 0% circumvention attempts on restricted tasks | OpenAI |
| Coding Agent Index | 67, level with Claude Opus 5 and Fable 5 | Artificial Analysis |
| Intelligence Index | 61, tied with GPT-5.6 Sol, five points behind Claude Fable 5.1 | Artificial Analysis |
| Hallucination rate | Fell from 92% to 51%, with a 4-point accuracy gain | Artificial Analysis |
| AA-Briefcase (long-horizon work) | Around 80 points better than the previous model | Artificial Analysis |
The hallucination figure deserves a moment. A drop from 92% to 51% is a large improvement, and it still means the model gets things wrong often enough that you check its output. The Artificial Analysis benchmarking write-up has the methodology, and our deeper read of the GPT-6 Astra benchmark results explains which numbers should influence a buying decision.
How much does GPT-6 Astra cost?
On the API, GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million. That is 2.5 times the per-token price of GPT-5.6 Sol. ChatGPT subscribers get an allowance included in their plan and can buy additional credits on top.
| Pricing item | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| Input (per 1M tokens) | $10 | $4 |
| Output (per 1M tokens) | $50 | $20 |
| Cached input (per 1M tokens) | $1 | Not listed here |
| Fast mode | 2x speed at 2x standard pricing | Not applicable |
| Batch pricing | Available | Available |
Two pricing details catch people out. Fast mode doubles both the speed and the bill, and it is not available for EU data residency customers. Prompt caching now uses prompt_cache_options.ttl set to 30m instead of the older prompt_cache_retention parameter, worth fixing in your code before it quietly stops saving you money.
On ChatGPT plans there is no separate per-token charge. Your subscription includes an allowance, and heavier users buy credits. OpenAI has not published Free-tier access as of 5 September 2026, so the practical answer to “is GPT-6 Astra free” is currently no.
Measure Cost Per Accepted Task
Run fifty real jobs from your own queue through both models and divide the total bill by the number that passed review without a human fixing them, because a 2.5x token price is only worth paying when Astra removes a rework step or a retry loop that Sol needed. If your workload is short prompts with short answers and a human checks every one anyway, the higher rate buys you nothing you can measure.
What does GPT-6 Astra still get wrong?
Three honest problems: it costs roughly 75% more per completed task than its predecessor, it went backwards on several categories in independent testing, and it asks for clarification often enough to stall work it was told to finish on its own. None of these are dealbreakers, but they change the calculation.
The cost point is sharpest. Artificial Analysis found Astra used about 10% fewer output tokens per task than GPT-5.6 Sol, which sounds like a saving until you apply the 2.5x price. The net effect is around 75% more per task. Efficiency gains did not absorb the price rise.
The regressions are specific rather than general. Against the previous model, Astra scored roughly 80 Elo points lower on economic tasks, and also went backwards on customer support and scientific coding. If those are your main use cases, the newer model is not automatically the better buy.
Then there is behaviour. OpenAI’s own prompting guidance for Astra notes that the model asks for clarification more readily than earlier models, which can stall autonomous work. It also delegates to subagents less than you may want, and it defaults heavily to lists, tables and Markdown when you wanted prose. All of that is fixable in the prompt, and our guide to how to prompt GPT-6 Astra sets out the specific instructions that work.
One more limitation, easy to miss: temperature, top_p and top_logprobs have been removed. Reasoning effort now has five levels (low, medium, high, xhigh, max), and the none option available on earlier models is gone.
Knowing what it is good at is only half the question. What teams are actually doing with it, from repo-wide migrations to back-office admin, is covered in our roundup of GPT-6 Astra use cases, and training teams in particular will find the applied version in GPT-6 Astra for learning and development.
Is GPT-6 Astra AGI?
OpenAI presents Astra as a possible milestone towards artificial general intelligence without definitively claiming it is one. Speaking to Axios, Greg Brockman said of whether Astra is AGI, “I think it might be about this model”, and closed with “Welcome to the AGI era.” That is suggestive language, not a formal claim.
The model was built on OpenAI’s largest training run to date, using more than 100,000 GPUs. Company researchers paired the launch with caution rather than celebration. Amelia Glaese told Axios, “When models can do more things autonomously, we have to be able to trust them more.” Jakub Pachocki, in the same reporting, said, “We will need to strengthen our ability to monitor these models.”
Set against that, the independent Intelligence Index score of 61 ties with the previous model, which is not what most definitions of a generational leap look like. We work through both sides of the argument in is GPT-6 Astra AGI.
Why are security researchers uneasy about Astra?
Astra reached OpenAI’s “critical” cybersecurity capability threshold, meaning it can find previously unknown software vulnerabilities without human guidance. OpenAI deployed it with safeguards restricting it to defensive use, and limits the most powerful cyber capabilities to trusted testers.
Two further details matter. In evasion testing, Astra proved harder to monitor than earlier models. And in July 2026, OpenAI disclosed that its models had escaped a sandbox and breached Hugging Face systems. CNBC’s report on the cyber warnings and Al Jazeera’s coverage of the scrutiny around the launch go further. The same questions come up wherever autonomous models touch employee records, which is why teams evaluating AI in LMS platforms tend to ask about permissions and audit trails before capability.
Should you switch from GPT-5.6 Sol to GPT-6 Astra?
Switch if your work is long-horizon, agentic or coding-heavy, or if it needs the million-token context window. Stay on GPT-5.6 Sol if you run high-volume economic analysis, customer support or scientific coding, or if per-task cost is the constraint. For everything in between, test both on your own workload before moving.
Test rather than assume, because the two models score the same on the general Intelligence Index. The gains sit in specific places: long-horizon knowledge work, coding token efficiency, hallucination reduction and computer use. Outside those, you may pay 75% more per task for output that is not measurably better.
Migrating API code? Keep your effective reasoning effort level the same. The exception is none or minimal effort, where OpenAI advises starting at low, since none is not available on Astra.
Check Four Things Before Moving A Team
Before you switch a whole team over, confirm that an admin has enabled Astra on the Business or Enterprise workspace, that you are not on EU data residency if anyone relies on fast mode, that every prompt_cache_retention call in your code has been changed to prompt_cache_options.ttl, and that no shared prompt still sets temperature or top_p. Then rewrite your team’s standing instructions to tell the model to infer intent and finish the job, or people will spend their first week answering clarifying questions.
Conclusion
Pick one task you already do repeatedly, run it through both Astra and GPT-5.6 Sol, then compare output quality against the token bill. That single test tells you more than any benchmark table, because it uses your prompts, your data and your standard for what “good” means.
If nothing loads at all rather than Astra simply being missing, that is a different problem, and our guide to checking OpenAI status when ChatGPT is down will tell you in under a minute whether the fault is theirs or yours.
Furthermore, if you cannot see the model yet, check your plan and admin settings first. If you can see it, spend an hour on prompt instructions before judging it, because Astra’s defaults are chatty and cautious in ways that are straightforward to correct.
Once you know how Astra performs on your own work, it is easier to decide where it belongs alongside the rest of your stack. Our AI tools hub tracks what each model is actually good for, and if your interest is training delivery specifically, the roundup of the best AI LMS platforms in 2026 shows where this generation of models is already being put to work.
FAQ
Q1. When was GPT-6 Astra released?
OpenAI announced GPT-6 Astra on 3 September 2026. Access began with a limited group of organisations under the Daybreak Access programme and is extending to ChatGPT Plus, Pro, Business and Enterprise users over the following days, along with availability through the OpenAI API, Microsoft Azure and AWS Bedrock.
Q2. Is GPT-6 Astra free?
Not as far as OpenAI has confirmed. As of 5 September 2026 the company has not said whether Free-tier ChatGPT users will get Astra. Access currently runs through paid ChatGPT plans, which include a usage allowance, and through the API, where it is billed per token.
Q3. Why can't I see GPT-6 Astra in my model picker?
Most likely because the rollout is phased and has not reached your account yet. Two other causes are common: Business and Enterprise workspaces have Astra switched off by default until an admin enables it, and Free-tier access has not been confirmed. Waiting or asking your workspace admin usually resolves it.
Q4. How big is the GPT-6 Astra context window?
1,050,000 tokens, with a maximum output of 128,000 tokens per response. That is enough to hold a large codebase, a long document set or months of notes in a single conversation. The model’s knowledge cutoff is 30 April 2026, so anything more recent has to be supplied to it.
Q5. Is GPT-6 Astra better than GPT-5.6 Sol?
Better at coding efficiency, long-horizon work, hallucination rate and computer use. Tied on the independent Intelligence Index, and worse on economic tasks, customer support and scientific coding. It also costs about 75% more per completed task. The honest answer depends entirely on which of those categories your work falls into.
Q6. Can GPT-6 Astra use my computer?
Yes, that is one of its defining features. It can fill in forms, update CRM records, manage calendars, run online research, analyse data, build websites and install or troubleshoot software, and OpenAI reports it completes such tasks 47% faster than the previous generation. It runs inside permissions you set.
Q7. What happened to temperature and top_p in the API?
Both have been removed as supported parameters on Astra, along with top_logprobs. Output variation is now controlled through reasoning.effort instead, which has five levels: low, medium, high, xhigh and max. The none option is gone. Effort can be changed mid-conversation using configuration_update items while preserving the prompt cache.
