Best AI Models for Reasoning and Math

As of July 2026, the strongest general reasoning models are Anthropic's Claude Opus 4.8 and Claude Fable 5, OpenAI's GPT-5.5, and Google's Gemini 3.1 Pro and Gemini 3 Pro in its Deep Think mode, which trade the top spot depending on the task ^[1][2][3]. Claude Fable 5 leads the composite Artificial Analysis Intelligence Index and posts the highest Humanity's Last Exam score; GPT-5.5 and Qwen3-Max-Thinking report perfect AIME math results; and Gemini 3 Deep Think holds the highest independently verified ARC-AGI-2 abstract-reasoning score ^[2][4][5][6]. For open weights, GLM-5.2 and DeepSeek V4 (V4-Pro) are the reasoning leaders ^[7][8]. Because AIME and MATH-500 are now saturated, the benchmarks that still separate the frontier are GPQA Diamond, Humanity's Last Exam (HLE), and ARC-AGI-2.

This page ranks models by measured reasoning and math performance. For how these "thinking" models actually work (chain-of-thought, test-time compute, verifier training), see the reasoning model concept page.

The reasoning leaderboard at a glance

Last verified: July 2026. AIME = AIME 2025, pass@1, no tools, unless marked with a dagger for AIME 2026 or "wt" for with tools. GPQA Diamond, HLE and ARC-AGI-2 are percentages. MATH-500 is saturated and has been dropped from 2026 launch reports (shown as n/r). ARC-AGI-2 values are the ARC Prize Foundation verified semi-private scores; the higher self-reported "ARC-AGI-2" figures some labs publish use a different public set and are discussed below. HLE is the no-tools score. "n/r" means not reported by a source we could verify (never a zero).

Scroll sideways for more →

Sources: AA Intelligence Index and independent evals ^[1][2][9]; Anthropic system cards for Opus 4.8/4.7 ^[3][10]; OpenAI and Vellum transcription for GPT-5.5 ^[5][11]; Google model cards for the Gemini line ^[6][12][13]; DeepSeek, Moonshot, Z.ai, Qwen, MiniMax model cards ^{[7][8][14][15][16]}; ARC Prize verified scores ^[4][17].

Which AI model is best for reasoning right now?

On the composite Artificial Analysis Intelligence Index (updated July 7, 2026), the order at the top is Claude Fable 5 first, Claude Opus 4.8 second, and GPT-5.5 third, with Anthropic holding two of the top three slots ^[1][2]. Claude Fable 5, the general-availability edition of Anthropic's "Mythos" class released on June 9, 2026, launched at number one and is the first model to clear 53% on Humanity's Last Exam, more than seven points ahead of the next model ^[2]. Opus 4.8 (released May 28, 2026) is the strongest widely deployed flagship, and GPT-5.5 (April 2026) is OpenAI's leading reasoning model and the top scorer on the agentic side of the index ^[1][5].

No single model wins every reasoning task, so the honest answer is a per-use-case ranking.

Which model is best for the hardest math?

AIME 2025 is effectively solved: GPT-5.5, GPT-5.2, Qwen3-Max-Thinking and (with tools) Grok 4 and Kimi K2 Thinking all report 100%, and most 2026 launches have simply stopped listing it ^{[5][14][18][19]}. The newest open models report AIME 2026 instead, where GLM-5.2 scores 99.2% and Kimi K2.6 scores 96.4%, again near the ceiling ^[7][8]. Because the exam is saturated, it no longer separates the leaders.

For math that still discriminates, the differentiators are harder olympiad sets and FrontierMath. GPT-5.5 leads here, reporting roughly 51.7% on FrontierMath tiers 1 to 3 and 35.4% on the research-level tier 4, and Gemini 3 Pro reaches 100% on AIME 2025 only when allowed to run code ^[5][12]. The verdict: for the hardest math, GPT-5.5 and Gemini 3 Pro in Deep Think mode are the strongest, with Qwen3-Max-Thinking the best proprietary Chinese option and GLM-5.2 the best open-weight math model.

Which model is best for graduate-level science?

Two benchmarks matter here: GPQA Diamond (Google-proof graduate science multiple choice) and HLE (thousands of expert-written frontier questions across dozens of fields). GPQA Diamond is now near saturation: the frontier clusters between 91% and 95%, well above the roughly 70% scored by human PhDs in their own field ^[17][20]. The current GPQA Diamond leaders are Gemini 3.1 Pro at 94.3%, Claude Opus 4.7 at 94.2%, Gemini 3 Deep Think at 93.8%, and Opus 4.8, GPT-5.5 and Grok 4.3 in a tight band near 93.5% ^{[3][5][12][13]}.

HLE is the tougher, less saturated science and knowledge test, and here Anthropic leads: Claude Fable 5 scores 53.0% and Claude Opus 4.8 scores 49.8% with no tools, ahead of Gemini 3 Deep Think at 48.4% and Gemini 3.1 Pro at 44.4% ^{[2][3][6][13]}. On Scale's independent no-tools harness, Gemini 3.1 Pro tops the public board at about 46.4% ^[9]. With tools enabled, Opus 4.8 rises to 57.9%, the highest verified with-tools result ^[3]. The verdict: for graduate science, Claude Opus 4.8 and Claude Fable 5 lead on HLE, while Gemini 3.1 Pro edges the GPQA Diamond ranking.

Which model is best for abstract and novel reasoning?

ARC-AGI-2 is the benchmark built to resist memorization: novel visual puzzles where humans average around 60% but 2025-era models scored under 20%. On the ARC Prize Foundation's verified semi-private evaluation, the top base model is Gemini 3 Pro in Deep Think mode at 45.1%, followed by Claude Opus 4.5 Thinking at 37.6%, Gemini 3 Flash at 33.6%, and Gemini 3 Pro base at 31.1%; Grok 4 reached 15.9% and GPT-5 reached 9.9% ^[4][10][17]. A refinement system from Poetiq built on Gemini 3 Pro is the only entry above 50%, at 54%, but it is a scaffolded system rather than a single model ^[10]. The verdict: for verified abstract reasoning, Gemini 3 Pro in Deep Think mode is the strongest single model.

An important caveat: OpenAI, Google and Anthropic have started publishing much higher "ARC-AGI-2" numbers for their 2026 flagships, such as GPT-5.5 at 85.0%, Gemini 3.1 Pro at 77.1% and Claude Opus 4.7 at 75.8% ^[5][13]. These use the public evaluation set and self-run harnesses, not the ARC Prize verified semi-private set, and they sit at roughly the same magnitude as ARC-AGI-1 scores, so they should not be compared directly with the 15% to 54% verified figures above ^[4][21]. Treat any single "ARC-AGI-2" percentage without noting which regime it comes from as unreliable.

Which is the best open-weight reasoning model?

Open weights have closed much of the gap. GLM-5.2 from Z.ai (Zhipu, MIT license, released June 2026) is the highest-scoring open model on the Artificial Analysis Intelligence Index and posts 91.2% GPQA Diamond, 99.2% AIME 2026, and 40.5% HLE (54.7% with tools) ^[8]. DeepSeek V4-Pro (MIT license, April 2026, a 1.6-trillion-parameter mixture-of-experts model with 49B active and 1M context) is close behind at 90.1% GPQA Diamond and 37.7% HLE, and far cheaper to run ^[7][22]. Kimi K2.6 from Moonshot AI scores 90.5% GPQA Diamond and a strong 54.0% HLE with tools ^[14]. The verdict: GLM-5.2 is the best open reasoning model overall, with DeepSeek V4-Pro the best open value and Kimi K2.6 the strongest tool-using open model. Note that Alibaba's Qwen3.7-Max and Qwen3-Max-Thinking are strong reasoners but are proprietary, API-only, and not open weight ^[15][16].

Which reasoning model is the best value?

Among open models, DeepSeek V4-Pro delivers frontier-adjacent reasoning at about $0.44 per million input and $0.87 per million output tokens, and its smaller V4-Flash sibling runs near $0.14 in and $0.28 out, the cheapest capable reasoner here ^[7][22]. MiniMax M2 (about $0.30 in, $1.20 out) is another low-cost open option, though its 78% AIME and GPQA place it a tier below ^[16]. Among proprietary models, Grok 4.3 from xAI is priced aggressively at $1.25 in and $2.50 out while scoring around 90% GPQA Diamond ^[23]. The verdict: DeepSeek V4-Pro is the best value for top-tier reasoning, and Grok 4.3 is the best-value proprietary option.

How should you read these scores?

Three cautions apply to every number above. First, saturation: AIME 2025 and MATH-500 are essentially solved and have been retired from most 2026 model cards, so a perfect AIME score no longer signals a leader. GPQA Diamond is close behind, with the frontier compressed into a two-point band above human-expert level ^[17][20]. HLE and ARC-AGI-2 are the remaining benchmarks with real spread. Second, tools versus no tools: with-tools scores (code execution, search) run 5 to 15 points higher than no-tools scores on HLE and AIME, so only compare like with like. Third, self-report versus independent: launch-day figures are provider self-reports; independent trackers such as Artificial Analysis, Scale and Epoch AI often measure a few points lower, and the ARC-AGI-2 public-set figures noted above are not the same measurement as ARC Prize verified scores ^{[4][9][20][21]}. Scores in this fast-moving field change monthly; this page was verified in July 2026.

References

1. Artificial Analysis, "LLM Leaderboard and Intelligence Index," artificialanalysis.ai/leaderboards/models (accessed July 2026). ↩
2. Artificial Analysis, "Claude Fable 5 tops the Intelligence Index," artificialanalysis.ai/models/claude-fable-5 and /articles/claude-fable-5-mythos-intelligence-index. ↩
3. Anthropic, "Introducing Claude Opus 4.8," anthropic.com/news/claude-opus-4-8 (May 28, 2026); Claude Opus 4.8 System Card. ↩
4. ARC Prize Foundation, verified ARC-AGI-2 semi-private results, arcprize.org/leaderboard and x.com/arcprize (Gemini 3 Deep Think 45.14%, Gemini 3 Pro 31.11%, Grok 4 15.9%, GPT-5 9.9%). ↩
5. OpenAI, "Introducing GPT-5.5," openai.com/index/introducing-gpt-5-5 (April 23, 2026). ↩
6. Google, "Gemini 3 Deep Think," blog.google/products/gemini/gemini-3-deep-think (HLE 41.0% at launch, 48.4% after Feb 2026 update). ↩
7. DeepSeek, "DeepSeek V4 Technical Report," arxiv.org/abs/2606.19348 and huggingface.co/deepseek-ai/DeepSeek-V4-Pro (MIT license). ↩
8. Z.ai (Zhipu), "GLM-5.2," z.ai/blog/glm-5.2 and huggingface.co/zai-org/GLM-5.2 (MIT license, June 2026). ↩
9. Scale AI, "Humanity's Last Exam leaderboard (SEAL)," labs.scale.com/leaderboard/humanitys_last_exam. ↩
10. ARC Prize Foundation, "ARC Prize 2025 Results and Analysis," arcprize.org/blog/arc-prize-2025-results-analysis (Opus 4.5 Thinking 37.6%, Poetiq system 54%). ↩
11. Vellum, "Everything you need to know about GPT-5.5," vellum.ai/blog/everything-you-need-to-know-about-gpt-5-5 (transcribes OpenAI launch table). ↩
12. Google DeepMind, "Gemini 3 Pro Model Card," storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf (AIME 95% no tools, GPQA Diamond 91.9%, ARC-AGI-2 31.1% verified, HLE 37.5%). ↩
13. Google, "Gemini 3.1 Pro," blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro (GPQA Diamond 94.3%, HLE 44.4%, self-reported ARC-AGI-2 77.1%). ↩
14. Moonshot AI, "Kimi K2.6," huggingface.co/moonshotai/Kimi-K2.6 (GPQA Diamond 90.5%, AIME 2026 96.4%, HLE 34.7% / 54.0% with tools). ↩
15. Qwen (Alibaba), "Qwen3-Max-Thinking," qwen.ai/blog?id=qwen3-max-thinking (AIME 2025 100% with tools, GPQA 87.4%). ↩
16. MiniMax, "MiniMax M2," huggingface.co/MiniMaxAI/MiniMax-M2; Qwen3.7-Max page, artificialanalysis.ai/models/qwen3-7-max. ↩
17. Epoch AI, "GPQA Diamond benchmark," epoch.ai/benchmarks/gpqa-diamond (human PhD baseline 69.7%; Grok 4 87%). ↩
18. OpenAI, "Introducing GPT-5.2," openai.com/index/introducing-gpt-5-2 (first 100% AIME 2025, no tools). ↩
19. LM Council, "AI Model Benchmarks (July 2026)," lmcouncil.ai/benchmarks (cross-model AIME and GPQA board; Claude Opus 4.8 AIME 98.3%). ↩
20. Artificial Analysis, "GPQA Diamond and Humanity's Last Exam evaluations," artificialanalysis.ai/evaluations/gpqa-diamond and /humanitys-last-exam. ↩
21. Imbue, "The evolution of ARC-AGI-2 (public versus semi-private)," imbue.com/blog/2026-02-27-arc-agi-2-evolution. ↩
22. DeepSeek, "API pricing," api-docs.deepseek.com/quick_start/pricing; NIST/CAISI, "Evaluation of DeepSeek V4-Pro," nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro. ↩
23. xAI, "Grok 4.3 model documentation," docs.x.ai/developers/models/grok-4.3 (pricing and release, April 30, 2026). ↩