OpenAI · released Sep 4, 2026
GPT-6 Astra
GPT-6 Astra is OpenAI's strongest thinker and the pick for hard math and science problems.
Our pick for hard problems8 weeks oldThinks before answering
- Input price
- $10 / M
- Output price
- $50 / M
- Context
- 1.1M tokens
- Longest reply
- 128K
- Reads
- text, image, pdf
- Knows up to
- 2026-04-30
Where it shines
- It leads HLE at 54.8 and ARC-AGI-2 at 95.
- It ties for the best FrontierMath score at 93.7.
- It ranks third on the agent leaderboard.
Where it struggles
- It costs $10 in, $50 out, five times GPT-6.1 Sol's price.
- Its text arena score of 1442 trails Opus 5.5 by 74 points.
Where you can use it
- OpenAI APIAPI$10 in / $50 out per million tokens.
- OpenRouterAPIListed as
openai/gpt-6-astra
How it scores on each job
Scores run 0 to 100 and blend every measure we have for that job. The number after the dot is its rank among the 33 models in the lineup.
The evidence
| Measure | Score | Rank in lineup | Source |
|---|---|---|---|
| Epoch Capabilities Index | 169 | #2 | Epoch AI |
| LMArena text | 1442 | #22 | LMArena |
| LMArena hard prompts | 1463 | #20 | LMArena |
| LMArena instruction following | 1455 | #19 | LMArena |
| LMArena multi-turn | 1449 | #20 | LMArena |
| LMArena creative writing | 1421 | #22 | LMArena |
| LMArena writing & language | 1440 | #20 | LMArena |
| LMArena coding | 1489 | #17 | LMArena |
| LMArena math | 1466 | #21 | LMArena |
| LMArena business & finance | 1438 | #22 | LMArena |
| LMArena legal & government | 1461 | #20 | LMArena |
| LMArena medicine | 1421 | #24 | LMArena |
| WebDev Arena | 1774 | #2 | LMArena |
| Vision Arena | 1280 | #13 | LMArena |
| Search Arena | 1234 | #4 | LMArena |
| Document Arena | 1468 | #6 | LMArena |
| Agent Arena | #3 | #3 | LMArena |
| GPQA Diamond | 96% | #2 | Epoch AI |
| Humanity's Last Exam | 55% | #1 | Epoch AI |
| FrontierMath | 94% | #1 | Epoch AI |
| SimpleQA Verified | 76% | #2 | Epoch AI |
| ARC-AGI-2 | 95% | #1 | Epoch AI |
Compare it with these
4 weeks old
OpenAI
GPT-6.1 Sol
GPT-6.1 Sol matches Astra on hard math for a fifth of the price.
Hard problems
- Price
- $2 in · $10 out
- Context
- 1.1M
5 weeks old
Anthropic
Claude Opus 5.5
Claude Opus 5.5 is the best all-round model and wins five jobs, from coding to reading images.
WritingCodingWebsites & apps
- Price
- $4 in · $20 out
- Context
- 1M
8 weeks old
Anthropic
Claude Fable 5.1
Claude Fable 5.1 reads long documents better than any other model and stays strong at most jobs.
Long documents
- Price
- $10 in · $50 out
- Context
- 1M