Claude Opus 4
- Made by
- Anthropic
- Released
- Announced in
- Introducing Claude 4anthropic.com
- Event
- Claude 4 (Opus 4 and Sonnet 4)
- API record
- /api/v1/models/claude-opus-4
Benchmark results
| Benchmark | Score | Source check |
|---|---|---|
| SWE-bench Verifiedno extended thinking; bash and file-editing tools; full 500 problems | 72.5% | CheckedA person checked this number against its source on 17 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced. |
| parallel test-time compute · parallel test-time compute | 79.4% | CheckedA person checked this number against its source on 17 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced.CorroboratedThe same number is printed in 1 other lab document: 1 |
| GPQA Diamondextended thinking up to 64K tokens | 79.6% | CheckedA person checked this number against its source on 17 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced. |
| no extended thinking | 74.9% | CheckedA person checked this number against its source on 17 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced.CorroboratedThe same number is printed in 1 other lab document: 1 |
| parallel test-time compute: best of several samples by an internal scoring model · parallel test-time compute | 83.3% | CheckedA person checked this number against its source on 17 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced. |
| AIME 2025No tools · extended thinking up to 64K tokens | 75.5% | CheckedA person checked this number against its source on 18 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced.CorroboratedThe same number is printed in 1 other lab document: 1 |
| extended thinking up to 64K tokens; parallel test-time compute with internal scoring · parallel test-time compute | 90.0% | CheckedA person checked this number against its source on 18 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced. |
| no extended thinking | 33.9% | CheckedA person checked this number against its source on 18 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced.CorroboratedThe same number is printed in 1 other lab document: 1 |
| Terminal-Bench 1.0no extended thinking; Claude Code harness | 43.2% | CheckedA person checked this number against its source on 18 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced. |
| Claude Code harness; parallel test-time compute with an internal scoring model · parallel test-time compute | 50.0% | CheckedA person checked this number against its source on 18 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced. |
| the same agent as the other labs' models, not Claude Code; no extended thinking | 39.2% | CheckedA person checked this number against its source on 18 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced. |
Sources
Reported by Anthropic
- Introducing Claude 4anthropic.com · 22 May 2025
Original image 
Our capture of the source
