Data source
Epoch AI Benchmarking Hub
Epoch AI runs and collects AI benchmark results. We keep its own SWE-bench Verified and MirrorCode results, and only scores and costs for results it collected from other projects.
Used to score: Cost of a hard task.
Provides measurements for the scored signals linked below.
Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from https://epoch.ai/benchmarks. Results collected from other projects keep their original terms, and we quote only top scores from those.
Visit the sourceUsed for: AI coding benchmarksLicense: CC BY 4.0
SWE-bench Verified, best83%
MirrorCode, best79%
DeepSWE, best74%
FrontierSWE, best66%
What the data shows
Computed from every record, as of 2026-04.
- The best SWE-bench Verified score is 83%, from claude-opus-4-7.
- Top scores on the other coding benchmarks here run from 55% to 79%, so one benchmark does not settle the question.
Best score by benchmark
| Benchmark | Best score | Model |
|---|---|---|
| SWE-bench Verified | 83% | claude-opus-4-7 |
| MirrorCode | 79% | claude-opus-5-5 |
| DeepSWE | 74% | gpt-6-astra |
| FrontierSWE | 66% | gpt-6-astra |
| FrontierCode | 55% | claude-opus-5-5 |
Limitations
- Benchmarks only cover the tasks they include, and models can be tuned for them.
- Each model is dated by its release, not when it was tested, and not every model is tested on every benchmark.