Data source
models.dev
models.dev is an open catalogue of AI models and the services that sell them, kept by the makers of OpenCode and its contributors. We use it to tell which models Epoch AI tested on SWE-bench Verified have open weights, then compare the best open and closed scores.
Provides measurements used by the linked signal.
Calculated from models.dev and its git history, as of 2026-11-01. MIT License.
Open models behind by4 months
Best open score79%
Best closed score83%
What the data shows
Computed from every record, as of 2026-11-01.
- glm-5.2, released in June 2026, scores 79% on SWE-bench Verified, which claude-opus-4-6 reached in February 2026.
- 1 of 35 results are left out, because models.dev doesn't list their model.
Each new best open-weight score
| Released | Open model | Score | Closed model that reached it first | Months behind |
|---|---|---|---|---|
| 2026-01 | kimi-k2.5 | 74% | claude-opus-4-5-20251101 (2025-11) | 2 |
| 2026-04 | deepseek-v4-pro | 78% | claude-opus-4-6 (2026-02) | 2 |
| 2026-06 | glm-5.2 | 79% | claude-opus-4-6 (2026-02) | 4 |
Limitations
- Its flags are edited by contributors and can be wrong.
- Models it doesn't list are left out of the comparison.