Best AI Models for Math

Mathematical-reasoning models ordered by the math pillar from the available published evaluations.

updated Oct 9, 2026 · 15 models ranked

How this list is ranked: Ranked by the math pillar using the snapshot’s benchmark weights. Check coverage and individual evaluations before treating a small lead as decisive.

Current inputs: frontiermath tier 1 3 (weight 1); frontiermath tier 4 (weight 1); frontiermath tier 4 v2 (weight 1); frontiermath tiers 1 3 v2 (weight 1); frontiermath vv2 tier 1 3 (weight 1); frontiermath vv2 tier 4 (weight 1); livebench math amps hard 2026_06_25 (weight 1); livebench math integrals with game 2026_06_25 (weight 1); livebench math math comp 2026_06_25 (weight 1); livebench math olympiad 2026_06_25 (weight 1); livebench math simplify 2026_06_25 (weight 1); math level 5 (weight 1); otis mock aime 2024 2025 (weight 0.5). Pillar scores can have different evidence coverage; review the individual results.

#	Model	Math pillar	SI Score	Confidence	Price in	Context
1	GPT-6.1 Sol OpenAI	93.0	73.0	100% confidence 100 percent, Full	$2.00	1.1M
2	GPT-6 Astra OpenAI	92.9	77.6	100% confidence 100 percent, Full	$10.00	1.1M
3	Claude Opus 5.5 Anthropic	91.5	78.4	100% confidence 100 percent, Full	$4.00	1M
4	Claude Fable 5.1 Anthropic	91.5	80.2	100% confidence 100 percent, Full	$10.00	1M
5	Gemini 3 Pro Preview Google provisional	91.4	58.6	68% confidence 68 percent, Medium	—	1M
6	GPT-6 Sol OpenAI	91.3	68.9	100% confidence 100 percent, Full	$2.00	1.1M
7	Qwen3.6 Max Preview Alibaba / Qwen provisional	91.1	66.4	64% confidence 64 percent, Medium	$1.30	262K
8	Claude Fable 5 Anthropic provisional	91.0	76.8	100% confidence 100 percent, Full	$10.00	1M
9	Claude Sonnet 5.5 Anthropic	90.5	69.4	93% confidence 93 percent, High	$2.00	1M
10	GPT-5.6 Sol OpenAI provisional	89.3	72.1	100% confidence 100 percent, Full	$4.00	1.1M
11	GPT OSS 120B OpenAI provisional	88.9	54.8	79% confidence 79 percent, Medium	—	131K
12	DeepSeek V4.1 Flash DeepSeek	88.0	66.2	69% confidence 69 percent, Medium	$0.15	1M
13	Claude Opus 5 Anthropic provisional	86.9	74.9	100% confidence 100 percent, Full	$5.00	1M
14	Muse Spark 1.2 Meta provisional	86.7	62.3	53% confidence 53 percent, Medium	$1.25	1M
15	DeepSeek-R1 DeepSeek provisional	86.6	47.9	88% confidence 88 percent, High	$0.55	128K

Task pages rank on one transparent metric, not the blended SI Score — see methodology for how pillars and confidence are computed. Missing values mean the source has not reported them.