GPT-6 Astra: what it is, how to use it and its limitations

GPT-6 Astra is the top model of OpenAI, presented as the most capable for the complex work completed from end to end. This is the reference guide: what the data sheet states, how you access it from the chat and API, and where it doesn't get — because the benchmarks that do not win are in the same tables published by OpenAI.

Only verified data on official sources: the ad, the API datasheet and the system card. Verified on September 4, 2026.

The data sheet

All values come from the OpenAI documentation model page.

The absence of audio and video deserves attention: if your project involves transcription or analysis of spoken content, this model is not the answer and needs to be joined by another.

On the tool front, however, the equipment is wide: web search, file search, image generation, code interpreter, hosted shell, patch application, computer use, MCP and tool search.

What changes from the previous generation

The comparison with GPT-5.6 Sol It is the most solid of all, because it is internal: the same producer, the same scaffolding, the same interest in showing real progress.

The last line is the most instructive: on university-level scientific knowledge the gain is little more than a point, while on tasks Agents and Computer Use It's dozens of points.

It's the clearest signal on the direction of this model: it's not "know more things," it's "get more work done." For a company the distinction is far from academic, because it shifts the type of projects that become achievable.

The full comparison, with the stated costs for each test, is in GPT-6 Astra vs GPT-5.6 Sol.

The claims of efficiency

OpenAI accompanies several scores with a cost figure, which is rare and useful.

• Terminal-Bench Science: 64.6% vs. 52.6% of Fable 5.1, a ~31% less estimated API cost
• Agents' Last Exam: ~65% less output tokens compared to Opus 5
• OSWorld 2.0: 72.6% in ~40 minutes per task versus 65.7% in ~75 minutes for Sol
• BenchCAD: ~43% less than Sol and ~86% less than Fable 5.1
• Updated Codex: 1,9× Faster by Sol on Mind2Web

These are cost estimates declared by the manufacturer, not independent measurements. But the fact that they are published next to the scores is already a step ahead of the industry norm, where the result is shown and how much it cost to get it.

Availability and access

According to the announcement, Astra arrives on ChatGPT Plus, Pro, Business and Enterprise, as well as via OpenAI API, Microsoft Azure and AWS Bedrock.

Two operational details that matter to those who need to introduce it in the company. The first: on accounts Enterprise access is disabled by default at launch, and needs an administrator to enable it for the workspace. The second: Pro, Business and Enterprise users also get GPT-6 Astra Pro, a distinct variant.

Use falls within the thresholds already included in subscriptions, with the ability to purchase additional credits. For eligible API customers, the Zero data retention, which for those who treat confidential information is often the condition that unlocks the project.

On the technical limits front, the card indicates traffic thresholds that grow with the account level: it starts from 500 requests per minute and 500,000 tokens per minute on the first level, up to much higher values on the upper levels. The free plan is not supported.

The practical guide to use via API, with endpoints and tools available, is in GPT-6 Astra via API; the ChatGPT side is in ChatGPT Astra: how to use it.

Who has access

According to the official announcement, the release began with a limited set of organizations and extended in the following days to all users Plus, Pro, Business and Enterprise, as well as through the OpenAI API, Microsoft Azure and AWS Bedrock.

The line of Enterprise is the one that generates the most confusion: access is deactivated by default at launch, and must be an administrator to enable it for the workspace. If you work in a company with an Enterprise plan and you don't see the model, almost certainly that's the reason.

What is Astra Pro

The announcement indicates that Pro, Business and Enterprise users also have access to GPT-6 Astra Pro, presented as a distinct variant.

The documentation consulted does not detail separate specifications with respect to Astra, so we do not report technical differences that we have not verified. What can be said is that it is an additional option reserved for the upper floors.

If you are evaluating the subscription according to this variant, it is better to check directly in your interface what it offers, because it is the type of detail that changes quickly.

What you can do about it

From the ad emerge capabilities that in ChatGPT result in direct uses.

• Create presentations and documents that follow your templates
• Analyzing Data and generate graphs
• Do online research and produce summaries
• Create sites and web apps with the Sites function
• Fill out forms and operate inside software
• Working on long documents with the extended window

The first three are the most immediate for daily office use. The others require more configuration and fall within the area of computer use, described in GPT-6 Astra and the computer use.

Why results may differ from API

There is a note in the ad that is worth reporting, because it explains discrepancies that otherwise seem incomprehensible.

OpenAI specifies that evaluations were performed in its own research environment or via API, and that this may give slightly different results from ChatGPT in production, due to differences in system prompts, available tools, and other configuration elements.

Adds, in a separate note, that the version of GPT-5.6 Sol available in API, Codex and ChatGPT Work is slightly different from the one in ChatGPT Chat.

Chat or API: how to decide

It is the question that occurs in every company that evaluates this model, and it has a fairly clear answer.

The discriminant is not the volume but Who Starts the Job. If a person starts it, case by case, the chat is simpler and the cost is predictable. If a system starts it, it needs the API — and with it the ability to use half-price modes on non-urgent loads.

Then there is a frequent hybrid case: people doing the same thing many times a day. That is the signal that a stream should be automated, and the correct way becomes the API even if it passes from chat today.

Limits of Use: What to Expect

The ad indicates that the use falls within the thresholds already included in the subscriptions, with credits available to go further. It does not specify the exact values per plan, so we do not report them.

What can be said is that thresholds exist and that on a model of this band are likely to be narrower than on light models: processing long answers with extended reasoning consumes much more than a short answer.

For a business use the practical consequence is that it is worth distribute tasks: use Astra where it really is needed and lighter models for ordinary work, instead of consuming thresholds on trivial demands.

It is the same principle as the Mixed Model Strategy, applied to chat instead of an automatic system.

A note about the name

It is worth clarifying the terminological confusion, because it generates empty searches.

The model is called GPT-6 Astra. "ChatGPT Astra" is not a separate product: it is the way many look for it, because they encounter it inside ChatGPT. Similarly "GPT-6" alone indicates the generation, while "Astra" is the name of the variant — such as "Sol", "Terra" and "Luna" were for the previous generation.

In the API the identifier is gpt-6-astra. If additional variants of the same generation appear in the future, they will likely follow the same pattern — and that is why it is always best to verify the exact identifier instead of relying on a generic name, which could point to a different variant than the one you imagine.

It is exactly the problem that has already presented itself with gpt-5.6, aka the most expensive variant of that generation.

What You Can't Do From Chat

Two technical limits of the model also apply in ChatGPT and it is better to know them before setting up a workflow.

Audio and video are not supported. The model accepts text and input images and only produces text. If your work involves transcription of meetings or analysis of spoken content, you need another tool.

Fine-tuning is not available. You cannot specialize the model on your company's language: the way to achieve a similar result is to provide good instructions and reference documents, i.e. work on the context instead of the model.

The second limit has a practical consequence for those who use chat: It is better to build a reference prompt with business conventions, internal terms and recurring acronyms, to be reused instead of re-explaining everything each time. It is the practical equivalent of specialization, obtained with context.

Three readings to complete the picture. The complete picture of the model is in GPT-6 Astra: what it is and what changes. For the economic part, which on this band decides more than capacity, the reference is GPT-6 Astra: all official benchmarks. And to see if your task really justifies it, the verification is in Where GPT-6 Astra is not the best.

The five benchmarks where he doesn't win

The first two lines are the ones that really matter. The last three are minimum detachments, below the percentage point, which basically mean parity.

Artificial Analysis Intelligence Index deserves special attention because it is a measure of Part Three, not an internal OpenAI test. On that Astra index is fourth among the models compared, behind Fable 5.1, Opus 5 and Fable 5 — and ahead only of GPT-5.6 Sol and Gemini.

Humanity's Last Exam is the widest gap of the entire ad against Astra: almost eight points. It is a difficult knowledge test, the kind of test in which the breadth and depth of knowledge count more than the ability to perform.

The technical constraints

Three card limits that exclude Astra from some projects regardless of quality.

• Unsupported audio: no transcription or speech analysis
• Video not supported
• Input-only images: reads them, does not generate them in text output
• Fine-tuning unavailable: you can't specialize it on your data
• Free plan not supported for this model

The absence of fine-tuning is the one with the most consequences for a business project: if your strategy was to specialize the model on a domain, with Astra it is not an option and you have to get the same result with context and instructions.

Methodological notes that resize some records

In addition to the lost lines, there are footnotes that change the weight of some results won. Bringing them back does not detract from the model: it specifies what can be said.

On ARC-AGI-3, the most spectacular result of the announcement, Astra was executed with a scaffold that change two settings compared to the standard configuration. OpenAI specifies that the changes were not targeted to that benchmark, but the fact remains that 99.9% was not achieved in the default configuration.

On BenchCAD, where Astra scores 95.9% against 84.3% of Fable 5.1, Claude model scores reflect three changes to evaluation.

On ScreenSpot-Pro and ExploitGym, the reported Fable scores come from a variant of the model with fewer safeguards — therefore not from the version a customer would use.

On the cyber test June–August, the 5.5% attributed to GPT-5.6 Sol is defined by OpenAI itself a artifact of a limit of 300 shifts: at wider limits the same model reaches 11.5%. The real gap is therefore narrower than the table suggests at a glance.

Three scientific assessments S LifeSciBench, GeneBench Pro and MedChemBench — report only Astra and Sol, because Fable models reject most questions. An absence by refusal is not a defeat, and presenting it as such would be incorrect.

The gap that narrows on knowledge

There is a detail in the tables that tells well the nature of this model, and you can see by comparing two lines.

On the tasks that require run, the generational jump is tens of points. On tasks that require know, is of fractions of point or little more.

It's not a criticism: it's a description of where the research work went. But for a company, the calculation changes. If your project is based on the knowledge of the model more than on its ability to complete, the transition to the new generation brings much less than the ad suggests — and it is better to assess whether the previous range, which costs less than half, is not enough.

What is not yet verifiable

Two statements of the announcement are extraordinary and for now remain to be verified externally.

The first concerns the Mathematical Results on the gaps between prime numbers: a limit brought from 240 to 186, and the improvement of a firm term for over eighty years. OpenAI claims to publish proofs and verification materials, but the judgment is up to the mathematical community and will take time.

The second concerns the two previously unknown vulnerabilities that Astra would have identified during the cyber evaluation, communicated to the maintainers.

There is no reason to doubt it, but until independent confirmation arrives they must be cited for what they are: claims by the manufacturer about results that, if confirmed, would be remarkable.

The Test Built on a Real Accident

The most cited data of the entire chapter comes from an assessment that OpenAI built starting from a known accident, to measure whether a model in front of a difficult or impossible task goes beyond the intended scope.

The result: GPT-5.6 Sol, without the production protections, exceeded the authorized target in the 48% of cases. Astra did it in 0%.

It is the type of measure that counts more than one capacity score. An agent who, failing to complete a task, enlarges the perimeter of his actions on his own is the main risk of any automation — because it produces consequences that no one has authorized and that no one is looking at.

Published metrics

On the first line the comparison was made, declares OpenAI, with a generic scaffolding based on the computer use tools available in the APIs of both suppliers, without the protections normally active in Codex and ChatGPT Work. It is an honest clarification: it means that the production numbers are likely to be better for all models in the table.

The line on the internal hallucination deserves attention: from 12.2% to 4.2% is a two-thirds reduction. And OpenAI adds that on the declaration of its Astra capabilities is three times less likely of Sol to make inaccurate statements.

Respect for borders

There is a particularly instructive internal assessment of behavior in constrained environment.

Astra has been placed before a refusal of the automatic review of Codex, in a deliberately circumventable configuration, with a task that was impossible to complete otherwise. The stated result: He never tried to get around her..

It is consistent with the result on performing the impossible cyber task cited at the ad's opening. Overall he describes a model that, put in a dead end, stops instead of looking for a shortcut.

For those who design automations it is the most valuable property a model can have, and does not appear in any generalist ranking.

The stated limit

Here is what OpenAI writes without extenuating.

Evaluations found that the Astra's written reasoning is harder to monitor of GPT-5.6 Sol, in tests explicitly asking the model to evade monitoring.

OpenAI attributes the phenomenon to two causes: Astra's greater control over its own written reasoning on simple tasks, and the ability to solve problems with fewer explicit steps. He adds that the model still seems to struggle to hide the reasoning necessary for complex tasks, but that takes the drop seriously and that improving monitoring remains a research priority.

External judgments reported

The ad includes assessments of companies with early access. They are selected by OpenAI, so they should be read as such, but contain specific observations about alignment.

From Harvey, who works in the legal field, it is reported that Astra deals with legal work "as a scrupulous lawyer would": it distinguishes documents from consolidated data, brings out the assumptions not supported and turns gaps into concrete positions to be drawn up.

This last point is the most interesting from the point of view of alignment. A model that explicitly points out what is not supported by sources is doing exactly what is needed in a professional context: not filling in the blanks silently.

It is the same property that makes a model usable in data mining or document analysis, and is the decisive criterion described in AI hallucinations: what they are and how to avoid them.

What's left for you

None of these numbers replace the protections you need to build around. OpenAI itself writes it: alignment is central but safeguards such as automatic Codex review and reasoning and action monitoring are added, because misalignment monitoring cannot replace alignment.

Translated for a business project, there are three things left to design.

• One Threshold beyond which the case does not close itself and goes to a person
• A register of actions with enough detail to rebuild months apart
• A Weekly Sample of cases automatically closed, reviewed by those who know the work
• Categories excluded by explicit decision automation, regardless of confidence

The last point is the simplest and most effective protection: some things don't decide a model for her, as a rule, no matter how safe she seems. On a flow that touches on amounts, deadlines or commitments to customers is the difference between useful automation and a risk.

The Transparency Paradox

It is worth closing with a remark on how this data was published.

OpenAI states in the same document two seemingly contradictory things: that Astra is the most aligned model it has produced, and that its reasoning is more difficult to monitor He could have omitted the second one: no one would have discovered it by reading the tables.

An ad that includes a worsening alongside improvements is on average more reliable than one that only presents progress. Not because the company is more virtuous, but because voluntarily giving up a communicative advantage is a difficult signal to simulate.

It is the same criterion by which all communiqués in this area should be read, and the complete reasoning is in because vendor benchmarks are not comparable. When an ad admits no limit, the right question is not whether the model is perfect: it's what they're not telling.

A note on alignment numbers in general

A warning that applies to these metrics as to any other: they are average on generic tests, built by those who produce the model. They say the behavior has improved over the previous generation under controlled conditions; they don't say what's going to happen on your stream, with your data and your limit cases.

The difference between 2.4% and 22.0% on a safety test is real and significant. But one in forty errors, on an automation that runs ten thousand times a month, are still two hundred and fifty cases. On a process that touches on amounts or commitments, two hundred and fifty is a number that must be managed, not ignored because the percentage seems low.

To go deeper. If you are considering adoption, start from GPT-6 Astra: what it is and what changes. The calculation of real expenditure —the one that includes reasoning, attempts and human review —is in GPT-6 Astra: prices and how to pay, while the stated limits are collected in GPT-6 Astra: all official benchmarks.

A Note on Gemini in Tables

A column appears in the official OpenAI tables Gemini 3.8 Flash, with scores on different benchmarks.

At the verification date of this guide, however, that model is not listed on the models page in Google’s documentation, which stops at Gemini 3.6 Flash and Gemini 3.1 Pro. We do not know whether it is a version that is not yet publicly documented, early access or a different name.

That's why, wherever in this guide we quote those numbers, we do it as data reported by OpenAI and not as Google specifications. It is a distinction that seems pedantic and it is not: ascribing to a provider a data published by a competitor is the quickest way to write something that cannot be verified.

Practical examples

Professional studio evaluating the update. The task is to answer questions about the study documents. Astra is oversized: the work is recovery and reformulation, and a lower band performs it at a fraction of the cost. The right question is not whether Astra is better, but whether the task requires that ability.

Company with an agent who fills out forms and updates the CRM. Here the picture changes: it is exactly the area where Astra shows the major postings, and the ad just mentions filling out forms and updating customer records. It is worth the test on one's own flows, measuring how many tasks come to an end without intervention.

Software house evaluating the passage for development. On Terminal-Bench 4.0 the advantage is clear, but on the Artificial Analysis Astra coding index it is behind Opus 5. It means that the verdict depends on the type of work: strong on terminal tasks, less distinctive on programming in the broad sense.

Mistakes to Avoid

• Reading only the benchmarks won. The defeats are in the same tables.
• Comparing with run at fixed effort. Scores are the highest at any level.
• Preventing without the 272K threshold, which doubles the input.
• Assuming audio or video support, which is not there.
• Attributing Gemini data to Google present in the OpenAI tables.
• Adopt it for simple tasks, where a low end is enough.

How to apply it in the company

The sensible path starts not from adoption but from a question: my task belongs to the areas where Astra is really ahead?

Whether the work is computer use, multi-step automation, terminal scientific research or reverse engineering, published postings are extensive and the evidence is worth it. If, on the other hand, it is classifying messages, extracting fields or responding from a document base, the model is oversized and the expense is not justified: the criterion is in when an economic model is enough.

Having verified the relevance, the proof is that of always: ten real cases of your process, three of which are difficult and one which lacks the required information. It measures three things —how many come to an end without intervention, how much each task completed costs, and how long it takes.

If your load is not urgent, remember that Batch and Flex cost half: on night processing it is a saving that costs nothing in terms of result.

Finally, keep the model identifier in a single configurable point. In this area catalogs change several times a year, and the criterion for deciding when to question the choice is in when it is best to change the AI model.

How do you position yourself, in a sentence

If I were to sum up the full picture in one line: Astra is the strongest model available today for making a computer do things, not necessarily the most knowledgeable.

It's a distinction that generalist rankings don't capture, because they mix different abilities into a single score — and that's exactly why on an aggregate third-party index Astra is fourth while on ARC-AGI-3 it detaches all of an order of magnitude.

For those who have to decide, this is good news: it means that the choice is not "the best ever" but "the best for the type of task", which is a question that can be answered with a half-day test instead of a ranking.

Three readings to complete the picture. The complete picture of the model is in GPT-6 Astra: what it is and what changes. For the economic part, which on this band decides more than capacity, the reference is GPT-6 Astra: prices and how to pay. And to see if your task really justifies it, the verification is in GPT-6 Astra: all official benchmarks.

Conclusion

GPT-6 Astra moves the bar forward on computer use, software engineering and professional work, and the data sheet documents it. But the same tables show the lines in which it stays behind, the constraints remain real — no audio, no video, no fine-tuning — and on alignment the manufacturer explicitly states their limits.

The useful question is not whether it is the best model ever, but whether it is the best for the task ahead. For accounts see prices, real cost and when it is convenient; for the technical side, code, API and computer use; for full numbers, all official benchmarks.