Time machine
Pick any day since 2023 to see the almanac as it stood then.
We date each score by its model’s release, so a model a board tested later shows up early. Since November 8th, 2026, we also record the day each new score appears.
November 21st, 2025
Gemini 3 Pro led both the capabilities index and Arena’s text leaderboard, among the 905 models out.
Top of the capabilities index
Epoch AI’s index as it stands today, among the models out by then.
- 1
Gemini 3 Pro153
- 2GPT-5 Pro151
- 3GPT-5.1151
- 4GPT-5150
- 5Claude Sonnet 4.5148
- 6o3-pro148
- 7o3147
- 8Grok 4147
- 9GPT-5 mini146
- 10Kimi K2 Thinking146
Arena’s text leaderboard
As Arena published it on November 20th, 2025.
- 1
Gemini 3 Pro1495
- 2Grok 4.1 thinking1481
- 3Grok 4.11462
- 4GPT-5.11454
- 5
Gemini 2.5 Pro1451
- 6Claude Sonnet 4.5 thinking1449
- 7Claude Opus 4.1 thinking1449
- 8Claude Sonnet 4.51444
- 9GPT-4.51442
- 10Claude Opus 4.11440
Watch a benchmark
Pick one to see its leaderboard on this day.
Try one of that day’s most-cited benchmarks.
Who led the most-cited benchmarks?
The 7 benchmarks that model releases cited most in the year before.
| Benchmark | Leader | Best | Led since |
|---|---|---|---|
| GPQA Diamond | No scores yet | ||
| AIME 2025 | Grok 4 Heavy | 100% | July 2025 |
| LiveCodeBench v6 | 87.6% | August 2025 | |
| MMLU-Pro | 87.2% | March 2025 | |
| IFEval | Claude 3.7 Sonnet | 93.2% | February 2025 |
| MMMU | GPT-5.1 | 85.4% | November 2025 |
| MMLU | o1 | 91.8% | December 2024 |
What did labs launch?
50 launch posts in the 90 days before.
- Nov 20, 2025
Olmo 3 32B ThinkLaunch post
- Nov 19, 2025Grok 4.1 FastLaunch post
- Nov 19, 2025
Cogito v2.1Launch post
- Nov 19, 2025GPT-5.1-Codex-MaxLaunch post
- Nov 18, 2025
Gemini 3 ProLaunch post
- Nov 17, 2025Grok 4.1Launch post
- Nov 13, 2025GPT-5.1Launch post
- Nov 11, 2025
ERNIE-4.5-VL-28B-A3B-ThinkingLaunch post
- Nov 6, 2025Kimi K2 ThinkingLaunch post
- Oct 31, 2025
LongCat-Flash-OmniLaunch post
- Oct 29, 2025Composer 1Launch post
- Oct 29, 2025SWE-1.5Launch post
And 38 more.
How much of the catalog was past 85% over time is now on How are evals changing?
Sources: Epoch Capabilities Index, Epoch AI (CC BY 4.0), Arena’s leaderboard dataset on Hugging Face (CC BY 4.0), labs’ launch posts, read by Benchmark Almanac and official leaderboards. Leaders and watched boards use the same source as each benchmark’s page.