AI benchmarks we track
Benchmark scores for the AI models in the Modelspike tracker, with the date each score was published and a link to its official source.
Last updated: Nov 18, 2026
The four benchmarks we track
Every model page records scores for four benchmarks that cover the main capability axes — coding, science reasoning, multimodal understanding and agentic execution:
- SWE-Bench Verified — agentic coding: resolving real GitHub issues with human-verified test cases.
- GPQA Diamond — graduate-level science questions, tested without tools.
- MMMU — multimodal understanding across charts, diagrams and documents.
- Terminal-Bench — terminal and agentic tasks executed end to end.
Current top scores
Among the models currently tracked, GPQA Diamond is the only one of the four with official published scores. Scores are not comparable across benchmarks — each measures a different capability mix.
| Model | GPQA Diamond (no tools) | As of | |
|---|---|---|---|
| 1 | GPT-6.2 Astra | 97.0% | Nov 17, 2026 |
| 2 | GPT-6.2 Sol | 96.0% | Nov 17, 2026 |
How we record scores
Every score comes from the model's official model card or release announcement. We record the score, the benchmark version, the date it was published and the source URL — model pages often carry additional official benchmarks beyond the four we track, such as Agents' Last Exam, FrontierCode 2.0 or BenchCAD, always with the same sourcing.
When a provider revises a number, the page updates and the 'as of' date moves with it. Nothing on a benchmark row is older than its stated date.
FAQ
Which benchmark should I trust most?
None of them alone. SWE-Bench Verified, GPQA Diamond, MMMU and Terminal-Bench measure different capability mixes — a model leading on one may trail on another. Compare within one benchmark, not across them.
Why do some models have scores on benchmarks not listed here?
Model cards often publish extra benchmarks — agentic suites like Agents' Last Exam, coding suites like FrontierCode 2.0, domain tests like BenchCAD. Those are recorded on each model's own page with sources and dates.