GPT-6 Astra
OpenAI's flagship. It ties Gemini 4 Argon at 53 on max or high settings and trails Opus 5.5 and Sonnet 5.5 at their peaks. It is token-efficient relative to Claude on measured tasks.
Every published effort
| Setting | Index | Per task | Tokens / sec |
|---|---|---|---|
| Max | 53 | $3.26 | 51 |
| Extra high | 52 | $2.31 | 48 |
| High | 51 | $1.73 | 45 |
| Medium | 50 | $1.54 | 45 |
| Low | 46 | $0.82 | 43 |
What this workload costs
List price for one call. Reasoning tokens count as output. A cache hit uses the lab’s cache rate when one is published.
Specs
- API id
- gpt-6-astra
- Released
- 3 Sep 2026
- Context
- 1.05M
- Max output
- 128k
- Input / 1M
- $10.00
- Output / 1M
- $50.00
- Cache read / 1M
- $1.00
- Knowledge
- 30 Apr 2026
- Inputs
- Text, Image
- Weights
- Closed
Above 272k input tokens the whole request is repriced, not just the overflow. Batch and Flex are half. Fast mode is double. Cache writes are $12.50.
Text and image in, text out, from the OpenAI model card.
Use it when
The hardest OpenAI jobs: long tool use, computer use, and work where you want the flagship specifically. High effort is close to max on the index and about half the measured task cost.
Leave it when
As the model every request hits. GPT-6.1 Sol is one index point behind at high effort and much cheaper per task. Prompts over 272k get the long-context multiplier on every token.
Tools named by the lab
- Web search
- File search
- Code interpreter
- Computer use
- MCP
Other independent scores
- Terminal-Bench 4.0: 59.6%. xhigh. Level with Opus 5.5 in that write-up. Sonnet 5.5 later scored 64%. Artificial Analysis, 22 Sep 2026.