GPT-6 Astra Is Here: What to Expect Compared with Fable 5.1 and GPT-5.6 SOL
OpenAI announced GPT-6 Astra, its new top-tier model, on September 3, 2026.
GPT-6 Astra focuses on completing difficult work that spans multiple stages, including computer use, web browsing, coding, and document creation.
The rollout has begun gradually with a limited number of organizations. OpenAI says Astra will expand to Plus, Pro, Business, Enterprise, and the API over the coming days. I have not been able to use Astra yet, so this article compares its published specifications and evaluations with Fable 5.1 and GPT-5.6 SOL, then outlines what I want to test once I gain access.
GPT-6 Astra is designed to complete difficult work
The move from GPT-5 to GPT-6 makes Astra OpenAI’s most capable model to date. What interests me most is not its ability to produce a single answer, but its ability to use tools while working through a long series of steps.
- Multi-stage work spanning computer use, browsing, and coding
- Taking new instructions into account and changing direction without losing sight of the original objective
- Continuing with other analysis or tasks while waiting for a tool to finish
- Creating documents, spreadsheets, and slides that match existing templates and writing styles
The API adds “Async tool calling.” While Astra waits for an external tool to return a result, it can call another tool or work on parts of the task that can be completed in advance.
“Mid-turn steering” also lets users send revised instructions while Astra is working, without forcing it to discard work it has already completed. For an AI that runs for a long time, being able to correct its course along the way is more important than writing a perfect initial prompt.
How much has Astra improved over GPT-5.6 SOL?
The first model to compare is GPT-5.6 SOL, which has been OpenAI’s flagship model until now.
Astra and SOL both support a 1.05-million-token context window and a maximum output of 128,000 tokens. However, Astra’s API input and output prices are both 2.5 times those of SOL.
| Specification | GPT-6 Astra | GPT-5.6 SOL |
|---|---|---|
| Input price | $10 | $4 |
| Cached input | $1 | $0.40 |
| Output price | $50 | $20 |
| Context window | 1.05 million tokens | 1.05 million tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
| Reasoning effort | low, medium, high, xhigh, max | none, low, medium, high, xhigh, max |
Prices are per 1 million tokens. GPT-5.6 SOL’s promotional prices of $4 for input and $20 for output will remain available until at least November 21, 2026. For both Astra and SOL, requests with more than 272,000 input tokens are charged at higher rates for the entire request.
The published evaluations show particularly large gains in automation and science-related terminal work.
| Evaluation | GPT-6 Astra | GPT-5.6 SOL |
|---|---|---|
| AutomationBench | 41.4% | 18.1% |
| Terminal-Bench 4.0 | 57.9% | 37.3% |
| DeepSWE v1.1 | 74.1% | 72.7% |
| Terminal-Bench Science 0.1 | 64.6% | 22.4% |
SOL remains close to Astra on DeepSWE, while the gap is much wider on AutomationBench, which measures the ability to carry out multi-step work. The deciding factor when choosing Astra will be whether its higher performance reduces trial and error and human revisions enough to lower the total time and cost required to finish the job. I discussed this way of thinking in “Choose AI models by cost per successful task, not token price.”
How does Astra compare with Claude Fable 5.1?
People who use Claude Code will naturally want to know how Astra compares with Anthropic’s Claude Fable 5.1.
Astra and Fable 5.1 have the same input and output prices, with little difference in their context windows or maximum output. However, Fable 5.1’s cached input price is one quarter of Astra’s, giving it an advantage for work that repeatedly refers to the same large set of materials.
| Specification | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| Input price | $10 | $10 |
| Cached input | $1 | $0.25 |
| Output price | $50 | $50 |
| Context window | 1.05 million tokens | 1 million tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
In the head-to-head evaluations published by OpenAI, Astra scored higher than Fable 5.1 in all four categories.
| Evaluation | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| AutomationBench | 41.4% | 31.4% |
| Terminal-Bench 4.0 | 57.9% | 55.8% |
| DeepSWE v1.1 | 74.1% | 67.4% |
| Terminal-Bench Science 0.1 | 64.6% | 52.6% |
The two models are close on Terminal-Bench 4.0, but Astra leads in automation and long-running software development.
These are evaluations published by OpenAI. Each score is the best result across reasoning-effort settings, and the results include research environments and API tests. Actual results in ChatGPT, Codex, and Claude Code will vary depending on the instructions, tools, and working environment.
What I expect from Astra: reaching the finish line without getting stuck
What I want most from Astra is not a single answer to a difficult question. I want it to take an assigned task, keep sight of the objective, use the tools it needs, and carry the work through to completion.
The first things I want to test are implementations that required repeated rework with SOL and large tasks where I am likely to change direction along the way. I want to see whether Astra can accept additional requests during a task while preserving both the work already completed and the original objective.
I also have high hopes for computer use. In OpenAI’s simulation using OSWorld 2.0, Astra achieved a higher score than SOL while reducing the time per task by about 47%. I want to see whether that speed and consistency are noticeable in work such as changing settings in a browser, checking the results, and trying again when necessary.
Another area I want to test is whether Astra can go beyond merely “creating” a document, spreadsheet, slide deck, or website and bring it to a genuinely usable state that matches an existing format. If it can check both appearance and content and take work close to submission or publication, its higher price may be worthwhile.
There is no need to replace Fable 5.1 or SOL immediately
Even with Astra’s arrival, there is no need to move every task to the most capable model.
For handling a high volume of everyday implementation and research, SOL is likely to remain practical: it has the same 1.05-million-token context window while its input and output prices are 60% lower. For difficult work that repeatedly refers to the same source materials, Fable 5.1 still has a clear advantage thanks to its lower cached input price.
For more about Fable 5.1, see “Claude Fable 5.1 released: what’s new and how it compares with GPT-5.6 SOL.”
References
- OpenAI, “GPT-6 Astra: A new generation of intelligence”
- OpenAI API, “GPT-6 Astra Model”
- OpenAI API, “Model guidance: GPT-6 Astra”
- OpenAI API, “GPT-5.6 Sol Model”
- Anthropic, “Claude Fable 5.1 and Claude Mythos 5.1”
Specifications, prices, and availability are current as of September 4, 2026. Performance figures are published by OpenAI, and actual results will vary depending on the environment and settings.
