AI Model Benchmarks
Track AI model performance across industry-standard benchmarks. Data sourced from official provider reports and third-party evaluations.
MMLU
Massive Multitask Language Understanding — general knowledge across 57 subjects. 2026 frontier values are estimates (providers no longer publish them).
HumanEval
Code generation benchmark — function-level completion from docstrings. 2026 frontier values are estimates.
GPQA
Graduate-level Google-Proof Q&A — expert-written science questions (Diamond subset)
SWE-bench
Real-world software engineering tasks from GitHub issues (Verified subset)
Score by Benchmark
Full Comparison Table
Sorted by average score. ★ = top score in column.
| Model | Provider | MMLU | HumanEval | GPQA | SWE-bench | Date |
|---|---|---|---|---|---|---|
| #1Claude Fable 5Suspended | Anthropic | ★93.0 | ★96.0 | 92.6 | ★95.0 | 2026-06 |
| Claude Opus 4.8 | Anthropic | 92.5 | 91.0 | 91.0 | 90.0 | 2026-05 |
| GPT-5.5 | OpenAI | 92.5 | 91.0 | 93.5 | 82.6 | 2026-04 |
| Gemini 3.1 Pro | 91.0 | 85.0 | ★94.3 | 80.6 | 2026-02 | |
| Muse Spark | Meta | 88.0 | 80.0 | 89.5 | 77.4 | 2026-04 |
| Grok 4.3 | xAI | 87.5 | 80.0 | 90.1 | 72.0 | 2026-04 |
| Claude Sonnet 4.6 | Anthropic | 89.0 | 84.0 | 80.0 | 72.0 | 2026-03 |
| DeepSeek-V3.2 | DeepSeek | 85.0 | 75.0 | 82.4 | 73.1 | 2025-12 |
| GPT-4o | OpenAI | 87.2 | 78.6 | 55.1 | 48.0 | 2024-05 |
Scores represent pass@1 or accuracy percentages as reported by providers or third-party evaluations. The dataset is refreshed weekly by an agent that web-searches provider announcements and leaderboards. GPQA (Diamond) and SWE-bench (Verified) figures are sourced; most 2026 frontier labs no longer publish classic MMLU / HumanEval headline numbers, so those columns for current models are directional estimates rather than provider claims. Models flagged “Suspended” may be temporarily unavailable — verify current status with the provider.