Leaders
We track which model held the best score on every benchmark since 2023, and for how long. A model keeps the top spot until another one beats it.
By default, each benchmark uses one source: its official leaderboard when that’s current, then Epoch AI’s own runs, then labs’ own numbers. Scores Epoch AI collected from lab reports count as claims.
Who’s at the frontier?
The top spot on Epoch AI’s Capabilities Index has changed hands 24 times since 2023. Claude Fable 5.2 from Anthropic holds it now. GPT-4 had the longest run at the top, which was 8 months.
Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)
Which lab leads most often?
OpenAI held the top spot 29% of the time across 348 benchmarks with at least three scores from two or more labs, using one source per benchmark. If a benchmark goes 45 days without a new score from another lab, we stop counting until one arrives. We stop for good once a model passes 85%.
Google DeepMind 12% 
Qwen 11% 
ByteDance Seed 5% xAI 3% Moonshot AI 3% Other labs 2% 
Meta 2% 
DeepSeek 2% Z.ai 2% 
NVIDIA 1% 
MiniMax 1% 
Shanghai AI Lab 1% 
Microsoft 1% 
Mistral AI 1% Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Who’s held on the longest?
GPT-5.4 has held number one on tau2-bench for 8 months, the longest run going today. We cover the 12 benchmarks most cited in model releases this past year, using one source per benchmark, and skip versions nobody has tested in six weeks.
Who’s held on the longest?
Phi-3.5-MoE held number one on TruthfulQA for 13 months, the longest run since 2023. We count only contested time, using one source per benchmark.
- 8 months GPT-5.4 on tau2-bench Mar 2026 to now
- 8 months Nemotron-Cascade 2 30B-A3B on IFBench Mar 2026 to now
- 7 months DeepSeek-V4-Pro on LiveCodeBench Apr 2026 to now
- 7 months DeepSeek-V4-Pro on LiveCodeBench v6 Apr 2026 to now
- 5 months Claude Fable 5 on BrowseComp Jun 2026 to now
- 4 months GPT-5.6 Sol on MMMU-Pro Jul 2026 to now
- 2 months GPT-6 Astra on Humanity's Last Exam Sep 2026 to now
- 2 months Mercury 2.5 on tau3-bench Sep 2026 to now
- 13 months Phi-3.5-MoE on TruthfulQA Aug 2024 to Jun 2026
- 13 months gpt-oss-120b on FORTRESS Aug 2025 to Sep 2026
- 12 months Gemini 2.5 Pro on LVBench Jul 2025 to Aug 2026
- 11 months o1 on AraGen Dec 2024 to Apr 2026
- 11 months Claude Opus 4 on Molecular Biology Capabilities Test May 2025 to Apr 2026
- 11 months GPT-4 on MedQA Jun 2023 to May 2024
- 10 months Grok 3 on TaxEval v2 Apr 2025 to Feb 2026
- 10 months Qwen2-72B on AlGhafa Jun 2024 to Aug 2025
Who finishes benchmarks off?
We call a benchmark saturated once a model scores past 85% on it, and that model gets the close-out. OpenAI closed out 49 of 142 benchmarks, the most of any lab.
- All areas
- Knowledge and language
- Math and science
- Coding
- Agents and tools
- Professional work
- Vision, audio and robotics
- Safety and security
- Other
Google DeepMind 17 Qwen 12 DeepSeek 5 Meta 4 Shanghai AI Lab 4 Other labs 4 ByteDance Seed 3 
Baidu 3 NVIDIA 2 Microsoft 2 
Meituan 2 xAI 2 
StepFun 1 
LG AI Research 1 Xiaomi Corp 1 
Arcee AI 1 MiniMax 1 
Tencent 1 Physical Intelligence 1 
Perplexity 1 
Ai2 1 Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
By category
Each category’s time at number one, split by lab. Click one to see its benchmarks.
Which lab leads most often?
OpenAI held the top spot 37% of the time across 343 benchmarks with at least three scores from two or more labs, using independent runs only. If a benchmark goes 45 days without a new score from another lab, we stop counting until one arrives. We stop for good once a model passes 85%.
Google DeepMind 14% Qwen 5% Other labs 3% Meta 2% xAI 2% DeepSeek 2% Shanghai AI Lab 2% Z.ai 1% Moonshot AI 1% ByteDance Seed 1% Sources: Official leaderboards and Epoch AI (CC BY 4.0)
Who’s held on the longest?
GPT-6 Astra has held number one on Humanity's Last Exam for 2 months, the longest run going today. We cover the 12 benchmarks most cited in model releases this past year, using independent runs only, and skip versions nobody has tested in six weeks.
Who’s held on the longest?
DeepSeek LLM 67B held number one on MuSR for 13 months, the longest run since 2023. We count only contested time, using independent runs only.
- 2 months GPT-6 Astra on Humanity's Last Exam Sep 2026 to now
- 2 months Claude Opus 5.5 on Terminal-Bench 4.0 Sep 2026 to now
- <1 month Gemini 4 Carbon on GPQA Diamond Oct 2026 to now
- 13 months DeepSeek LLM 67B on MuSR Nov 2023 to Jan 2025
- 13 months gpt-oss-120b on FORTRESS Aug 2025 to Sep 2026
- 11 months o1 on AraGen Dec 2024 to Apr 2026
- 11 months Claude Opus 4 on Molecular Biology Capabilities Test May 2025 to Apr 2026
- 11 months GPT-4 on MedQA Jun 2023 to May 2024
- 10 months Grok 3 on TaxEval v2 Apr 2025 to Feb 2026
- 10 months Qwen2-72B on AlGhafa Jun 2024 to Aug 2025
- 9 months Gemini 1.5 Pro on Video-MME May 2024 to Jun 2025
Who finishes benchmarks off?
We call a benchmark saturated once a model scores past 85% on it, and that model gets the close-out. OpenAI closed out 42 of 93 benchmarks, the most of any lab.
- All areas
- Knowledge and language
- Math and science
- Coding
- Agents and tools
- Professional work
- Vision, audio and robotics
- Safety and security
- Other
Google DeepMind 14 Qwen 6 Shanghai AI Lab 4 Meta 2 Xiaomi Corp 1 Z.ai 1 xAI 1 Writer 1 
Cohere 1 Sources: Official leaderboards and Epoch AI (CC BY 4.0)
By category
Each category’s time at number one, split by lab. Click one to see its benchmarks.
Which lab leads most often?
OpenAI held the top spot 32% of the time across 352 benchmarks with at least three scores from two or more labs, with claimed scores included. If a benchmark goes 45 days without a new score from another lab, we stop counting until one arrives. We stop for good once a model passes 85%.
Google DeepMind 12% Qwen 9% ByteDance Seed 3% Other labs 3% xAI 2% Meta 2% DeepSeek 2% Moonshot AI 2% Z.ai 2% Shanghai AI Lab 1% NVIDIA 1% Microsoft 1% MiniMax 1% Mistral AI 1% 
Amazon 1% Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Who’s held on the longest?
Gemini 3.1 Pro has held number one on MMLU-Pro for 9 months, the longest run going today. We cover the 12 benchmarks most cited in model releases this past year, with claimed scores included, and skip versions nobody has tested in six weeks.
Who’s held on the longest?
GPT-4o held number one on BigCodeBench for 18 months, the longest run since 2023. We count only contested time, using claimed scores included.
- 9 months Gemini 3.1 Pro on MMLU-Pro Feb 2026 to now
- 8 months GPT-5.4 on tau2-bench Mar 2026 to now
- 8 months Nemotron-Cascade 2 30B-A3B on IFBench Mar 2026 to now
- 7 months GPT-5.5 on AIME 2026 Apr 2026 to now
- 7 months DeepSeek-V4-Pro on LiveCodeBench Apr 2026 to now
- 7 months DeepSeek-V4-Pro on LiveCodeBench v6 Apr 2026 to now
- 5 months Claude Fable 5 on BrowseComp Jun 2026 to now
- 4 months GPT-5.6 Sol on MMMU-Pro Jul 2026 to now
- 18 months GPT-4o on BigCodeBench Jun 2024 to Apr 2026
- 18 months o3-mini on LiveBench Jan 2025 to Nov 2026
- 13 months DeepSeek LLM 67B on MuSR Nov 2023 to Jan 2025
- 13 months Phi-3.5-MoE on TruthfulQA Aug 2024 to Jun 2026
- 13 months gpt-oss-120b on FORTRESS Aug 2025 to Sep 2026
- 12 months Gemini 2.5 Pro on LVBench Jul 2025 to Aug 2026
- 12 months Gemini 1.5 Pro on Video-MME May 2024 to May 2025
- 11 months o1 on AraGen Dec 2024 to Apr 2026
Who finishes benchmarks off?
We call a benchmark saturated once a model scores past 85% on it, and that model gets the close-out. OpenAI closed out 54 of 147 benchmarks, the most of any lab.
- All areas
- Knowledge and language
- Math and science
- Coding
- Agents and tools
- Professional work
- Vision, audio and robotics
- Safety and security
- Other
Google DeepMind 21 Qwen 13 Meta 5 Shanghai AI Lab 4 Baidu 3 ByteDance Seed 3 DeepSeek 3 Other labs 2 StepFun 1 LG AI Research 1 NVIDIA 1 Microsoft 1 Xiaomi Corp 1 Meituan 1 MiniMax 1 xAI 1 Physical Intelligence 1 Perplexity 1 Cohere 1 Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
By category
Each category’s time at number one, split by lab. Click one to see its benchmarks.
Google DeepMind 0 Meta 0 StepFun 0 OpenAI 0 LG AI Research 0 Qwen 0 NVIDIA 0 Microsoft 0 Xiaomi Corp 0 Arcee AI 0 ByteDance Seed 0 Baidu 0 Meituan 0 DeepSeek 0 MiniMax 0 xAI 0 Tencent 0 Physical Intelligence 0 Perplexity 0 Shanghai AI Lab 0 Ai2 0 Other labs 0 Inception Labs 0 Moonshot AI 0 
Ant Group 0 Amazon 0 
Deep Cogito 0 
Stanford University 0 
INF Technology (infly) 0 Z.ai 0 Writer 0 Cohere 0 Mistral AI 0 Gemini 3 Deep Think Llama 3.1-405B STEP3-VL-10B GPT-4 Mar 2023 K-EXAONE 2.0 Qwen2.5-72B Nemotron 3 Ultra 550B-A55B Base MAI-Thinking-1 MiMo-V2.5 Trinity Large Thinking Seed1.5-VL ERNIE-4.5-21B-A3B-Thinking LongCat-Flash-Thinking DeepSeek-R1 MiniMax-M2 Grok-3 mini Hunyuan-T1 π0 Sonar Pro internlm2-20b-reward Tulu 3 405B PaLM 2-L Mercury 2.5 Kimi K3 Ring-2.5-1T Amazon Nova Act Cogito v2 Preview 671B MoE OpenVLA-OFT INF-ORM-Llama3.1-70B GLM-4.6 Palmyra X4 Command R (03-2024) Mistral Large 4 (Le Chonk, open weights)