Best AI for Math: The Arena Math Leaderboard
claude-opus-5-max tops Arena math, but 14 models from 5 labs are statistically tied with it. 4 of them are lighter "Flash" models.
The signal
- LaunchesClaude Opus 5 (24 Jul 2026): claude-opus-5-max leads math on 904 votes.
- LaunchesGemini 3.7 Flash (13 Aug 2026) and GLM-5.3-Flash (26 Aug 2026) both sit inside the leader's interval.
- ArenaThe median top-10 math model has 1,355 votes, against 25,382 on the overall board.
14 models are statistically tied for #1 in Arena math, spanning 49 points. 4 are Flash models from Google and Z.ai.
Top 18 models in the Arena math category, 13 Sept 2026. Bar = Arena points above the #50 math model (deepseek-v4-pro-high-preview, 1457).
Arena math is a category of the Arena text board: blind human votes on math prompts. Voters judge which answer they prefer; Arena does not check the answer against a key. "Statistically tied" means the model's 95% upper bound reaches the leader's lower bound (1518). "Flash" is taken from the model name; it marks a lab's lighter line, not a measured price. Latest snapshot on or before 15 Sept 2026.
Show the data (18 rows)
| Rank | Model | Lab | Rating | 95% CI | Votes | Tied with #1 |
|---|---|---|---|---|---|---|
| 1 | claude-opus-5-max | Anthropic | 1537 | 1518–1557 | 904 | yes |
| 2 | claude-opus-5-high | Anthropic | 1534 | 1520–1549 | 1,866 | yes |
| 3 | gemini-3.7-flash-high | 1525 | 1492–1558 | 310 | yes | |
| 4 | claude-fable-5.1-max | Anthropic | 1522 | 1485–1558 | 243 | yes |
| 5 | claude-fable-5 | Anthropic | 1522 | 1506–1537 | 1,467 | yes |
| 6 | claude-opus-4-6-high | Anthropic | 1516 | 1506–1526 | 3,695 | yes |
| 7 | claude-opus-4-6 | Anthropic | 1512 | 1502–1522 | 4,110 | yes |
| 8 | glm-5.3-flash | Z.ai | 1511 | 1483–1539 | 441 | yes |
| 9 | gemini-3.6-flash-high | 1506 | 1489–1523 | 1,242 | yes | |
| 10 | gemini-3.5-flash-high | 1505 | 1491–1519 | 1,867 | yes | |
| 11 | qwen3.8-max | Alibaba | 1500 | 1479–1521 | 769 | yes |
| 12 | glm-5.3-max | Z.ai | 1498 | 1471–1525 | 430 | yes |
| 13 | claude-opus-4-7-high | Anthropic | 1498 | 1487–1509 | 3,030 | no |
| 14 | kimi-k3-max | Moonshot | 1494 | 1474–1514 | 884 | no |
| 15 | qwen3.7-max-preview | Alibaba | 1493 | 1454–1533 | 218 | yes |
| 16 | gpt-5.4-high | OpenAI | 1490 | 1479–1501 | 3,179 | no |
| 17 | claude-opus-4-7 | Anthropic | 1489 | 1477–1500 | 3,144 | no |
| 18 | muse-spark-1.3-max | Meta | 1489 | 1450–1527 | 207 | yes |
The Mode read
Google and Z.ai shipped lighter Flash models in August, and both now sit near the top of the Arena math category. So what is the best AI for math, and does the leaderboard know?
claude-opus-5-max is #1 at 1537. But the math board has thin evidence: the median top-10 model has 1,355 votes and an interval of ±16 points, against ±5 on the overall board. As a result, 14 models from Anthropic, Google, Z.ai, Alibaba and Meta cannot be separated from the leader. The best Flash model is #3. OpenAI's best math model is #16.
When 14 models tie, a math product cannot win on the model. It wins on what surrounds the answer: the right next problem, the steps shown the way a teacher shows them, and access to the students and engineers who need it. A cheaper model that ties for first is a margin advantage, not a product.
14 models tie for the top in math. The tie goes to whoever already has the students.
What we would back: Math and quantitative tools that reach users through a channel they already own, such as schools, exam prep or engineering teams.