How good is AI getting?
Five of the hardest tests for AI, side by side. Compare how fast each is being solved, sort them, or open one for its full history and every model’s score.
Best score today
Each line is the best score any model had reached on that test by the end of the quarter.
- GPQA Diamond96%
- SWE-bench Verified84%
- FrontierMath94%
- Humanity’s Last Exam55%
- ARC-AGI-295%
Hover or tap the chart to step through quarters. Tap a name above to hide or show its line.
| Quarter | GPQA Diamond | SWE-bench Verified | FrontierMath | Humanity’s Last Exam | ARC-AGI-2 |
|---|---|---|---|---|---|
| Q1 2023 | 36% | No data | No data | No data | No data |
| Q2 2023 | 36% | No data | No data | No data | No data |
| Q3 2023 | 36% | No data | No data | No data | No data |
| Q4 2023 | 42% | No data | No data | No data | No data |
| Q1 2024 | 47% | No data | 0% | No data | No data |
| Q2 2024 | 54% | No data | 1% | No data | No data |
| Q3 2024 | 62% | No data | 1% | 5% | 1% |
| Q4 2024 | 77% | 31% | 15% | 8% | 1% |
| Q1 2025 | 84% | 61% | 19% | 18% | 3% |
| Q2 2025 | 85% | 71% | 36% | 22% | 9% |
| Q3 2025 | 87% | 74% | 55% | 25% | 16% |
| Q4 2025 | 93% | 77% | 74% | 38% | 54% |
| Q1 2026 | 95% | 79% | 83% | 46% | 85% |
| Q2 2026 | 95% | 84% | 88% | 46% | 89% |
| Q3 2026 | 96% | 84% | 94% | 55% | 95% |
| Q4 2026 | 96% | 84% | 94% | 55% | 95% |
The tests
- GPQA DiamondBest score today 96%+3 pts in 12 monthsRoom left Nearly maxed outRecord GPT-6 Astra (max), since Q3 2026Humans Experts in their own field · 81%Graduate-level biology, chemistry and physics questions that are hard to look up.
- SWE-bench VerifiedBest score today 84%+7 pts in 12 monthsRoom left Closing inRecord Claude Opus 4.7 (max), since Q2 2026Fixing real bugs in open-source Python projects so the tests pass.
- FrontierMathBest score today 94%+20 pts in 12 monthsRoom left Nearly maxed outRecord GPT-6.1 Sol (max), since Q3 2026Unpublished research-level maths problems with answers a computer can check.
- Humanity’s Last ExamBest score today 55%+17 pts in 12 monthsRoom left Closing inRecord GPT-6 Astra (unknown thinking), since Q3 2026Thousands of expert questions across 100+ subjects, chosen to stump AI.
- ARC-AGI-2Best score today 95%+41 pts in 12 monthsRoom left Nearly maxed outRecord GPT-6 Astra (max), since Q3 2026Visual grid puzzles that are easy for people and new to every model.
Benchmark data from Epoch AI (CC BY 4.0) and official leaderboards. Updated weekly, last on . Download CSV Methodology