Best AI for maths: all 43 ranked models
Ordered by hard maths problems. Score shown: how many very hard maths problems it solved. Higher is better, and neighbouring models are often too close to separate: check the range before reading much into a one-place gap.
This is one test, not a recommendation. Our pick for work through maths also looks at school-competition maths and maths votes, and flags close calls. See the options worth considering.
| # | Model | Position | Score | Rating | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra Good-enough alternativeOpenAI | Ranked 1st of 43 | 93.7%standard error ±1.4 points | Top tier | Source for GPT-6 Astra work through maths score (epoch) |
| 1 | GPT-6.1 Sol Best-supported choiceOpenAI | Ranked 1st of 43 | 93.7%standard error ±1.4 points | Top tier | Source for GPT-6.1 Sol work through maths score (epoch) |
| 3 | Claude Opus 5.5 Good-enough alternativeAnthropic | Ranked 3rd of 43 | 91.2%standard error ±1.7 points | Top tier | Source for Claude Opus 5.5 work through maths score (epoch) |
| 4 | Claude Fable 5.1Anthropic | Ranked 4th of 43 | 90.2%standard error ±1.8 points | Top tier | Source for Claude Fable 5.1 work through maths score (epoch) |
| 5 | GPT-6 SolOpenAI | Ranked 5th of 43 | 89.8%standard error ±1.8 points | Top tier | Source for GPT-6 Sol work through maths score (epoch) |
| 6 | GPT-5.6 SolOpenAI | Ranked 6th of 43 | 89.1%standard error ±1.8 points | Top tier | Source for GPT-5.6 Sol work through maths score (epoch) |
| 7 | Claude Sonnet 5.5Anthropic | Ranked 7th of 43 | 88.8%standard error ±1.9 points | Top tier | Source for Claude Sonnet 5.5 work through maths score (epoch) |
| 8 | Claude Fable 5Anthropic | Ranked 8th of 43 | 87.0%standard error ±2.0 points | Top tier | Source for Claude Fable 5 work through maths score (epoch) |
| 9 | Gemini 4 ArgonGoogle DeepMind | Ranked 9th of 43 | 86.0%standard error ±2.1 points | Top tier | Source for Gemini 4 Argon work through maths score (epoch) |
| 9 | GPT-5.6 TerraOpenAI | Ranked 9th of 43 | 86.0%standard error ±2.1 points | Top tier | Source for GPT-5.6 Terra work through maths score (epoch) |
| 11 | Claude Opus 5Anthropic | Ranked 11th of 43 | 85.6%standard error ±2.1 points | Top tier | Source for Claude Opus 5 work through maths score (epoch) |
| 12 | GPT-5.5OpenAI | Ranked 12th of 43 | 85.3%standard error ±2.1 points | Strong | Source for GPT-5.5 work through maths score (epoch) |
| 13 | GPT-5.6 LunaOpenAI | Ranked 13th of 43 | 82.1%standard error ±2.3 points | Strong | Source for GPT-5.6 Luna work through maths score (epoch) |
| 14 | Claude Opus 4.8Anthropic | Ranked 14th of 43 | 80.0%standard error ±2.4 points | Strong | Source for Claude Opus 4.8 work through maths score (epoch) |
| 15 | GPT-6 LunaOpenAI | Ranked 15th of 43 | 78.9%standard error ±2.4 points | Strong | Source for GPT-6 Luna work through maths score (epoch) |
Show the other 28 ranked models
| # | Model | Position | Score | Rating | Source |
|---|---|---|---|---|---|
| 16 | Claude Haiku 5.5Anthropic | Ranked 16th of 43 | 75.1%standard error ±2.6 points | Strong | Source for Claude Haiku 5.5 work through maths score (epoch) |
| 17 | Qwen3.8-MaxAlibaba | Ranked 17th of 43 | 74.7%standard error ±2.6 points | Strong | Source for Qwen3.8-Max work through maths score (epoch) |
| 18 | Muse Spark 1.3Meta | Ranked 18th of 43 | 74.4%standard error ±2.6 points | Strong | Source for Muse Spark 1.3 work through maths score (epoch) |
| 19 | Kimi K3 Best open-weights option Open weightsMoonshot | Ranked 19th of 43 | 72.2%standard error ±2.7 points | Strong | Source for Kimi K3 work through maths score (epoch) |
| 20 | Gemini 3.7 FlashGoogle DeepMind | Ranked 20th of 43 | 71.6%standard error ±2.7 points | Strong | Source for Gemini 3.7 Flash work through maths score (epoch) |
| 21 | Claude Opus 4.7Anthropic | Ranked 21st of 43 | 70.2%standard error ±2.7 points | Strong | Source for Claude Opus 4.7 work through maths score (epoch) |
| 22 | GLM-5.3 Open weightsZ.ai (Zhipu AI) | Ranked 22nd of 43 | 68.8%standard error ±2.7 points | Strong | Source for GLM-5.3 work through maths score (epoch) |
| 23 | Gemini 3.8 FlashGoogle DeepMind | Ranked 23rd of 43 | 68.4%standard error ±2.8 points | Middle | Source for Gemini 3.8 Flash work through maths score (epoch) |
| 24 | Claude Opus 4.6Anthropic | Ranked 24th of 43 | 66.0%standard error ±2.8 points | Middle | Source for Claude Opus 4.6 work through maths score (epoch) |
| 24 | Grok 4.6xAI | Ranked 24th of 43 | 66.0%standard error ±2.8 points | Middle | Source for Grok 4.6 work through maths score (epoch) |
| 26 | Claude Sonnet 5Anthropic | Ranked 26th of 43 | 65.6%standard error ±2.8 points | Middle | Source for Claude Sonnet 5 work through maths score (epoch) |
| 27 | DeepSeek-V4-Pro (High, 20260813) Open weightsDeepSeek | Ranked 27th of 43 | 64.6%standard error ±2.8 points | Middle | Source for DeepSeek-V4-Pro (High, 20260813) work through maths score (epoch) |
| 28 | Gemini 3.5 FlashGoogle DeepMind | Ranked 28th of 43 | 62.8%standard error ±2.9 points | Middle | Source for Gemini 3.5 Flash work through maths score (epoch) |
| 29 | GLM-5.2 Open weightsZ.ai (Zhipu AI) | Ranked 29th of 43 | 59.2%standard error ±3.0 points | Middle | Source for GLM-5.2 work through maths score (epoch) |
| 30 | Gemini 3.6 FlashGoogle DeepMind | Ranked 30th of 43 | 58.9%standard error ±2.9 points | Middle | Source for Gemini 3.6 Flash work through maths score (epoch) |
| 31 | Mistral Large 4Mistral AI | Ranked 31st of 43 | 58.6%standard error ±2.9 points | Middle | Source for Mistral Large 4 work through maths score (epoch) |
| 32 | Grok 4.5xAI | Ranked 32nd of 43 | 57.2%standard error ±2.9 points | Middle | Source for Grok 4.5 work through maths score (epoch) |
| 32 | Kimi K2.6 Open weightsMoonshot | Ranked 32nd of 43 | 57.2%standard error ±2.9 points | Middle | Source for Kimi K2.6 work through maths score (epoch) |
| 34 | GLM-5.3-Flash Open weightsZ.ai (Zhipu AI) | Ranked 34th of 43 | 55.8%standard error ±2.9 points | Behind the leaders | Source for GLM-5.3-Flash work through maths score (epoch) |
| 35 | Grok 4.7xAI | Ranked 35th of 43 | 53.0%standard error ±3.0 points | Behind the leaders | Source for Grok 4.7 work through maths score (epoch) |
| 36 | Inkling-Small Open weightsThinking Machines | Ranked 36th of 43 | 46.3%standard error ±3.0 points | Behind the leaders | Source for Inkling-Small work through maths score (epoch) |
| 37 | DeepSeek-V4-Pro Open weightsDeepSeek | Ranked 37th of 43 | 45.3%standard error ±3.0 points | Behind the leaders | Source for DeepSeek-V4-Pro work through maths score (epoch) |
| 38 | Qwen3.7-PlusAlibaba | Ranked 38th of 43 | 34.4%standard error ±2.8 points | Behind the leaders | Source for Qwen3.7-Plus work through maths score (epoch) |
| 39 | Qwen3.6-PlusAlibaba | Ranked 39th of 43 | 32.3%standard error ±2.8 points | Behind the leaders | Source for Qwen3.6-Plus work through maths score (epoch) |
| 40 | Qwen3.5-397B-A17B Open weightsAlibaba | Ranked 40th of 43 | 31.2%standard error ±2.7 points | Behind the leaders | Source for Qwen3.5-397B-A17B work through maths score (epoch) |
| 41 | GPT-5.5 InstantOpenAI | Ranked 41st of 43 | 26.3%standard error ±2.6 points | Behind the leaders | Source for GPT-5.5 Instant work through maths score (epoch) |
| 42 | Gemini 3.5 Flash-LiteGoogle DeepMind | Ranked 42nd of 43 | 26.0%standard error ±2.6 points | Behind the leaders | Source for Gemini 3.5 Flash-Lite work through maths score (epoch) |
| 43 | GLM-5.1 Open weightsZ.ai (Zhipu AI) | Ranked 43rd of 43 | 24.9%standard error ±2.6 points | Behind the leaders | Source for GLM-5.1 work through maths score (epoch) |
No result for this test
Not enough data for hard maths problems: amazon-nova-experimental-chat (26-02-10), Claude Opus 4.5 (20251101), Claude Sonnet 4.6, DeepSeek-V4-Pro Preview, DeepSeek-V4.1-Flash-20260910, Dola Seed 2.0 Pro, ERNIE 5.0 (0110), ERNIE 5.1, Gemini 2.5 Pro, Gemini 3 Flash, Gemini 3 Pro, Gemini 3.1 Pro Preview, Gemma 4 31B, GLM-4.7, GLM-5, GPT-5.1, GPT-5.4, Grok 4.20 Beta (Reasoning, 0309), Grok 4.20 Beta 1, Grok 4.20 Multi-Agent Beta (0309), Grok 4.3, Kimi K2.5, MiMo-V2.5-Pro, MiMo-V2.6-Flash, MiMo-V2.6-Pro, Muse Spark, Muse Spark 1.1, Muse Spark 1.2, Muse Wonder, NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4), Qwen3.5-Flash, Qwen3.5-Max-Preview, Qwen3.6-Max-Preview and Qwen3.7-Max-Preview. That is a gap in the evidence, not a low score.
Test: FrontierMath Tiers 1-3 (v2). Source: epoch. Data last changed 29 Oct 2026. How we test.
Want a straight answer for work through maths?
See the options the evidence supports, with limits and verified prices.