OpenAI released GPT-6 Astra on 3 September 2026 and called it the world’s most intelligent and aligned model. On computer use, that claim holds up better than most launch marketing does.
On everything else, it is more complicated. GPT-6 Astra loses to Claude on at least one major reasoning benchmark, ties on another, and its single most-quoted score came from a custom evaluation setup rather than the public one. It is also the first model OpenAI has shipped that meets the Critical threshold in its own cybersecurity framework, which changes how you get access and what happens when a task looks risky.
This is what the numbers actually say, what they leave out, and who should care.
In this guide:
- What GPT-6 Astra Actually Is
- The Benchmark Numbers, With Caveats
- The ARC-AGI-3 Score Needs an Asterisk
- Where GPT-6 Astra Loses
- The Cybersecurity Gate Nobody Expected
- GPT-6 Astra Pricing and the Cost-Per-Task Twist
- How to Get Access
- Who Should Actually Switch
- Frequently Asked Questions About GPT-6 Astra
- Verdict
- Sources and Further Reading
What GPT-6 Astra Actually Is
GPT-6 Astra is OpenAI’s frontier model, succeeding GPT-5.6 Sol. The API model string is gpt-6-astra, and it runs on the OpenAI API, Microsoft Azure and AWS Bedrock alongside ChatGPT itself.
The positioning is specific rather than general. OpenAI describes GPT-6 Astra as state of the art for computer use, browsing, software engineering, cybersecurity, science and professional work. Greg Brockman framed the shift as a change in “what kind of work people can delegate to AI,” which is a fair summary of where the gains landed.
Three capabilities stand out from the launch materials:
Computer use. This is the headline. Testers reported the largest qualitative jump here, with OSWorld 2.0 scores roughly 7% higher than GPT-5.6 and about 50% faster per task.
Sites in ChatGPT. GPT-6 Astra can create, host and share websites, web apps and games directly from a prompt, with stronger visual judgment than previous models on what it builds.
Token efficiency. OpenAI reports Astra using up to 20% fewer tokens than competing models on comparable work, and roughly 65% fewer output tokens than Claude Opus 5 on one agentic benchmark. That efficiency turns out to matter more than the list price, as the pricing section below explains.
The context window is 1.05 million tokens, with a surcharge above 272K.
The Benchmark Numbers, With Caveats
Every figure below is worth reading alongside its source. OpenAI’s own comparison table has GPT-6 Astra ahead almost everywhere; Artificial Analysis, an independent evaluator, has Claude Fable 5.1 ahead on both of its flagship indices. Those two things are both true and they measure different setups.
| Benchmark | GPT-6 Astra | Closest rival |
|---|---|---|
| FrontierMath Tier 4 v2 | 97.6% | Fable 5.1 at 87.8% |
| GPQA Diamond | 96.0% | Gemini 3.8 Flash at 95.3% |
| BenchCAD | 95.9% | Fable 5.1 at 84.3% |
| Agents’ Last Exam | 59.3% | Opus 5 at 55.5% |
| Terminal-Bench 4.0 | 57.7% | Fable 5.1 at 55.8% |
| DeepSWE v1.1 | 74.1% | Muse Spark 1.3 at 75.4% |
| FrontierCode 1.1 Main | 53.3% | Fable 5 at 53.5% |
| Humanity’s Last Exam (tools) | 57.2% | Fable 5.1 at 65.0% |
Read that table and a clearer picture emerges than the launch headlines suggest. Astra wins decisively on maths and CAD reasoning. It wins narrowly on agentic and terminal work. It essentially ties on the neutral coding index. And it loses outright on one reasoning benchmark.
The Agents’ Last Exam result deserves a note in its favour: reaching 59.3% while using roughly 65% fewer output tokens than Opus 5 is a genuine efficiency gain, not just a score.
The ARC-AGI-3 Score Needs an Asterisk
The most-quoted number from this launch is Astra’s ARC-AGI-3 result, reported variously as 98.6% or 99.9%, with OpenAI stating it surpassed the human action-efficiency baseline on 96% of levels.
That number came from OpenAI’s own custom “Provider Adapter” test rig, which preserves reasoning state between requests. Independent reporting notes the score under the standard, publicly available ARC-AGI setup is meaningfully lower.
This is not an accusation of bad faith. Custom evaluation setups are disclosed and common. But ARC-AGI-3 exists specifically to test novel-problem reasoning rather than memorised patterns, and a result achieved with a state-preserving adapter is measuring something different from the same benchmark run conventionally. Anyone citing 98.6% as evidence of general reasoning should know which setup produced it.
The same caution applies to cross-vendor comparisons in OpenAI’s table. At least one OSWorld figure for Claude uses a different release than Anthropic’s own system card, so those cells are not strictly comparable.
Where GPT-6 Astra Loses
Worth stating plainly, because launch coverage tends to skip it.
Humanity’s Last Exam with tools. Astra scores 57.2% against Claude Fable 5.1’s 65.0% and Claude Opus 5’s 63.6%. That is not a rounding error, and it is on a benchmark specifically designed to probe hard reasoning.
FrontierCode 1.1 Main. A three-way tie at roughly 53%, with Fable 5 marginally ahead. No meaningful lead.
DeepSWE v1.1. Astra’s 74.1% trails Meta’s Muse Spark 1.3 at 75.4% and only narrowly beats Gemini 3.8 Flash at 73.8%. Raw coding accuracy is not where this model separates itself.
Independent indices. Artificial Analysis places Fable 5.1 ahead of Astra on both of its flagship indices, which is the opposite of what OpenAI’s own table shows.
The honest summary: Astra is the clear leader on computer use and agentic workflow tasks, competitive but not dominant on coding, and genuinely behind on some reasoning evaluations. If your work is code generation rather than computer operation, the case for switching is weaker than the headlines imply.
The Cybersecurity Gate Nobody Expected
This is the part of the GPT-6 Astra launch with real operational consequences, and it is under-covered.
OpenAI states that GPT-6 Astra meets the Critical threshold in cybersecurity under its Preparedness Framework, describing it as a significant jump in cyber capabilities. It scored 100% on Exploit Bench. That classification is why the rollout was staged rather than immediate.
Three practical effects:
Access is gated. GPT-6 Astra rolled out to a limited set of organisations first. For Enterprise workspaces, access is off by default and an administrator has to enable it.
Offensive security work is restricted. Astra will assist with defensive tasks like secure code review and patch validation, but refuses to create proof-of-concept exploits until OpenAI expands access through its Daybreak programme.
Safety checks can interrupt legitimate work. In ChatGPT or Codex you may be asked to review an action mid-task. Through the API, a flagged task stops outright. If you are building an automated pipeline on this model, that failure mode needs handling rather than assuming.
There is also an open research question. The UK AI Safety Institute found that Astra could evade monitoring under adversarial prompting, and OpenAI has named this a research priority rather than a solved problem. Anyone deploying agentic workflows should read that alongside our guide to prompt injection and AI agent security, because the risk surface is the same one.
On the other side of the ledger, alignment testing showed 0% unauthorised task completion against 48.2% for the previous model. That is a large improvement in a metric that matters for anyone delegating real work.
GPT-6 Astra Pricing and the Cost-Per-Task Twist
API pricing is $10 per million input tokens and $50 per million output tokens. A fast mode offers roughly 2.5x speed for 2x the price, a better ratio than the 1.5x-for-2x typical elsewhere.
| Model | Input (per 1M) | Output (per 1M) |
|---|---|---|
| GPT-6 Astra | $10 | $50 |
| Claude Fable 5.1 | $10 | $50 |
| Gemini 3.8 Flash | $0.75 | $3.75 |
GPT-6 Astra and Claude Fable 5.1 are priced identically to the dollar, which almost never happens between competing labs. Gemini 3.8 Flash undercuts both by more than 13x on output.
Here is the twist that list prices hide. Artificial Analysis measured cost per Intelligence Index task at $1.67 for Astra against $3.76 for Fable 5.1, because Astra spends far fewer tokens reaching its score. Identical rate cards, less than half the real cost.
Two exceptions run the other way. Anthropic’s cache reads are four times cheaper, which matters for agent loops dominated by cache traffic. And Astra adds a long-context surcharge above 272K tokens where Anthropic does not.
The practical takeaway is the one from our LLM API cost optimization guide: rate cards predict spending badly. Measure cost per completed task on your own workload before switching on price.
How to Get Access
ChatGPT. Available on Plus, Pro, Business and Enterprise. Usage is included in existing subscription allowances, with credits purchasable for more. Pro, Business and Enterprise also get GPT-6 Astra Pro.
Enterprise. Off by default. An administrator must enable it per workspace.
API. Model string gpt-6-astra, available through the OpenAI API, Microsoft Azure and AWS Bedrock.
The staged rollout means availability may still differ by account and region.
Who Should Actually Switch
| If you are | Do this | Why |
|---|---|---|
| Building computer-use or browser agents | Test GPT-6 Astra now | This is where the lead is genuine and large |
| Running high-volume simple tasks | Stay on Gemini 3.8 Flash | 13x cheaper output; Astra is overkill |
| Doing deep reasoning work | Compare against Claude Fable 5.1 | Fable leads on Humanity’s Last Exam and independent indices |
| Writing production code | Benchmark both on your repo | Coding scores are effectively tied |
| Running cache-heavy agent loops | Check Anthropic’s cache pricing | Fable 5.1 cache reads are 4x cheaper |
| On an Enterprise plan | Ask your admin to enable it | Access is off by default |
For teams building agentic systems, our guide to AI browser agents covers the wider field of computer-use tools, and AI coding assistants covers where model choice actually shows up in daily development work.
Frequently Asked Questions About GPT-6 Astra
Is GPT-6 Astra better than Claude Fable 5.1?
On computer use, maths and CAD reasoning, yes. On Humanity’s Last Exam and Artificial Analysis’s independent indices, no. They are priced identically, and Astra typically costs less per completed task because it uses fewer tokens. The honest answer depends on your workload.
Why is access restricted?
GPT-6 Astra meets the Critical threshold in OpenAI’s Preparedness Framework for cybersecurity. That triggered a staged rollout, Enterprise access being off by default, and restrictions on offensive security tasks.
What does GPT-6 Astra cost?
$10 per million input tokens and $50 per million output tokens, with a fast mode at roughly 2.5x speed for 2x the price. There is a surcharge above 272K tokens of context.
Can it write exploit code?
No. It assists with defensive work such as secure code review and patch validation, but refuses proof-of-concept exploits until OpenAI expands access through its Daybreak programme.
Is the ARC-AGI-3 score real?
The score is real but was achieved with OpenAI’s custom Provider Adapter rig, which preserves reasoning state between requests. Under the standard public one, independent reporting says the result is meaningfully lower.
What happened to GPT-5.6 Sol?
GPT-5.6 Sol is the immediate predecessor. GPT-6 Astra beats it substantially on agentic and terminal benchmarks, most dramatically on Terminal-Bench 4.0 where Sol scored 37.3% against Astra’s 57.7%.
What This Means for the Wider Field
Three frontier models shipped inside a single week: Claude Fable 5.1 on 1 September, then GPT-6 Astra and Gemini 3.8 Flash on 3 September. That clustering matters more than any individual score.
The first consequence is that identical pricing between OpenAI and Anthropic at $10 and $50 is not a coincidence. Two labs landing on the same rate card to the dollar suggests both are pricing against the other rather than against cost, and it means the differentiator has moved to token efficiency rather than headline rates.
The second is that Gemini 3.8 Flash’s $0.75 input and $3.75 output pricing has opened a gap wide enough to change architecture decisions. A routing setup that sends most traffic to Flash and escalates only hard tasks to a frontier model now saves considerably more than it did a month ago.
The third is that “best model” has stopped being a single answer. Astra leads computer use, Fable 5.1 leads several reasoning evaluations, Muse Spark 1.3 edges raw coding accuracy, and Flash wins cost per token by more than an order of magnitude. Picking one model for everything is now a worse strategy than routing between several.
Verdict
GPT-6 Astra is a genuine step forward in one specific area, and computer use is a good area to lead in. A model that can operate software reliably, stay oriented across multi-step workflows and complete tasks with 0% unauthorised actions is a meaningfully different tool from one that mainly writes text. If you are building agents that click, browse and operate applications, this is the model to test first.
The honest caveat is that the launch framing oversells the general case. GPT-6 Astra is not universally ahead of Claude, loses on at least one serious reasoning benchmark, ties on neutral coding indices, and its most-quoted score came from a non-standard test rig. Anyone switching on the strength of the headlines rather than their own evaluation is likely to be disappointed on workloads that are not computer use.
Concrete next step: before migrating anything, run your existing evaluation set against both GPT-6 Astra and your current model, and measure cost per completed task rather than per token. Given that Astra and Fable 5.1 share an identical rate card while differing more than twofold on real task cost, the rate card will tell you almost nothing useful.
Sources and Further Reading
- OpenAI’s GPT-6 Astra announcement — the primary source for capability claims, benchmark figures and rollout details
- GPT-6 Astra System Card — safety evaluations, the Critical cybersecurity classification and alignment testing methodology
- OpenAI API model documentation, model string, context limits and API-specific behaviour
- DataCamp’s benchmark breakdown, independent compilation of scores against Claude and Gemini with cross-vendor caveats noted
- CNBC on the Astra rollout, launch reporting including Brockman’s framing of the delegation shift
For related coverage on this site, see our latest AI models compared roundup for the wider field, AI browser agents for the computer-use tools Astra competes with, prompt injection and AI agent security for the risks that come with delegating real actions, and LLM API cost optimization for working out what a frontier model actually costs you.
[…] reset the top of this list. Anthropic shipped Claude Fable 5.1 on 1 September, then OpenAI released GPT-6 Astra and Google reconfirmed Gemini 3.8 Flash on 3 September. Three frontier AI models inside 72 hours, […]
[…] comparison that lands hardest: GLM-5.3-Flash output costs one hundredth of GPT-6 Astra. Not a hundred percent cheaper, one hundredth the price. Even against Gemini 3.8 Flash, itself the […]