A few weeks ago we wrote about Qwen 3.8 bringing open-weight models into the top tier. On the closed-model side, OpenAI has now answered: GPT-6 Astra opened to a limited set of organizations on 3 September 2026 and to broad availability the next day.
OpenAI positions the model as the successor to GPT-5.6 Sol and its new flagship. In this article we look at what Astra really offers, where its benchmark table shines, and where it needs a careful read.
What Sets GPT-6 Astra Apart?
The areas OpenAI emphasizes are computer use, software engineering, science, professional work and cybersecurity. In practice this means Astra is less a chat model than a model designed to run multi-step work end to end.
Key technical specs:
- Context window: 1,050,000 tokens, with up to 128,000 output tokens per response.
- Knowledge cutoff: 30 April 2026.
- Long-context performance: 96.3% retrieval on MRCR v2 8-needle in the 512K to 1M token range.
- Access: ChatGPT Plus, Pro, Business and Enterprise plans, the OpenAI API, Microsoft Azure and AWS Bedrock. On Enterprise accounts it is off by default and enabled per workspace.
For computer use, OpenAI states that Astra spends roughly 47% less time per task on OSWorld 2.0 than GPT-5.6 Sol. In agent scenarios, speed is as much a cost line as accuracy.
What the Benchmarks Show
The table below contains the rows from OpenAI's published results where competitor data exists:
| Benchmark | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5 |
|---|---|---|---|
| OSWorld 2.0 (computer use) | 72.6 | Not reported | 70.2 |
| Terminal-Bench 4.0 | 57.7 | 55.8 | 52.3 |
| DeepSWE v1.1 | 74.1 | 69.9 | Not reported |
| GPQA Diamond | 96.0 | 93.7 | 93.7 |
| FrontierMath Tier 4 | 97.6 | 87.8 | 73.2 |
| Humanity's Last Exam (with tools) | 57.2 | 65.0 | 63.6 |
The table says two things at once. Astra leads on coding, terminal and math-heavy tests. But it is not a clean sweep: on Humanity's Last Exam with tools it trails both Fable 5.1 and Opus 5. On some tests, such as FrontierCode 1.1, the three models are almost tied.
The Caveats Behind the Numbers
Some headline figures take on a different meaning once read together with their footnotes.
The 99.9% on ARC-AGI-3
This is the most shared number. But the result was achieved with a special stateful test harness. With standard, stateless API calls the same test yields between 17% and 63% depending on the chosen reasoning level. So the performance you get through the API in your own application may be very different from the headline.
OSWorld results
OSWorld 2.0 scores depend on latency simulation. There is no guarantee of the same result on a real corporate desktop with real network and application latency.
The whole table is from OpenAI
All the numbers above are OpenAI's own measurements, and some rows have no competitor data at all. The picture may shift as independent evaluations arrive.
Cybersecurity: At the "Critical" Threshold for the First Time
This may be the most important thing that separates Astra from earlier models. Under OpenAI's own risk framework, the model reaches the "Critical" threshold for cybersecurity capabilities. In the company's definition, this means that with the right tools and access, the model can find previously unknown flaws in well-protected systems and develop ways to exploit them.
That has practical consequences:
- Secure code review and patch generation are available, but producing working exploit code (proof of concept) is refused at launch.
- Advanced capabilities such as exploit validation, malware analysis and detection engineering open up through controlled access via OpenAI's Daybreak program.
- Extra safety checks can sometimes stop legitimate defensive work too. In ChatGPT and Codex the model may ask for confirmation, while in the API the task may simply stop.
For security teams this is a dilemma: the model's defensive value is high, and exactly for that reason access is restricted.
Pricing and Who It Fits
Astra is not a cheap model:
| Item | Price (per 1 million tokens) |
|---|---|
| Input | 10.00 dollars |
| Cached input | 1.00 dollar |
| Output | 50.00 dollars |
There is also a fast mode that runs roughly 2.5 times faster and is billed at twice the rate. For comparison: Qwen3.8-Max is offered through its API at 2 dollars input and 6 dollars output per million tokens.
The price also defines where Astra makes sense. It is a strong candidate for multi-step agent tasks, work on complex codebases and professional work where mistakes are expensive. For routine work such as bulk text generation, classification or simple summarization, the cost is hard to justify; those jobs can be done far more economically with smaller models.
Known Limitations
Reasoning monitorability has regressed. Astra produces shorter, more controlled reasoning chains. That brings efficiency, but makes it harder to monitor how the model thinks. In the UK AI Security Institute (UK AISI) evaluation, monitor evasion behavior was observed under adversarial prompting.
Cost versus capacity. The high output price can grow the bill quickly in agent loops that produce long responses. It should not go into production without a budget cap.
Safety checks can interrupt workflows. Teams working in security, infrastructure and system administration in particular should expect legitimate tasks to be stopped unexpectedly.
Conclusion
GPT-6 Astra pushes the frontier of closed models forward on agent tasks and coding. But the benchmark table does not show one-sided dominance, some headline figures depend on special conditions, and its price clearly makes it a premium option.
Our advice is the same as with every major model announcement: measure the model on your own workload, against your own cost and latency targets. A short comparison for the same task between Astra, the Claude family and an open-weight alternative like Qwen 3.8 usually says more than the headlines. We run model selection and agent architecture work under our AI Solutions service. We covered the impact of autonomous models on software development in our Claude Fable 5 article.
