Who held number one
We track which model had the best score on each benchmark at every point since 2023. A model takes the top spot on the day its score beats every earlier one, and keeps it until another model beats it. A tie leaves the holder in place.
The top spot on the capability index has changed hands 21 times since 2023
GPT-6 Astra, from OpenAI, holds it now. The longest run at the top since 2023 was 8 months.
OpenAI held number one 33% of the contested time
Each lab’s share of the days its models held number one, across 333 benchmarks with enough scores, lab claims included. We count only contested time. A benchmark’s clock stops when its best score passes 85%, or at its newest score once 45 days go by without another. A new version doesn’t stop the old one’s clock while people still test it.
Google DeepMind 14% 
Qwen 11% 
ByteDance Seed 4% xAI 3% Other labs 2% 
Meta 2% Z.ai 2% Moonshot AI 1% 
DeepSeek 1% 
Shanghai AI Lab 1% 
Mistral AI 1% 
MiniMax 1% Physical Intelligence 1% 
NVIDIA 1% The longest runs at the top
Lab claims included, counted from January 2023 and only while the benchmark was contested.
- 26 months GPT-4o on OR-Bench May 2024 to Jul 2026
- 19 months o3-mini on LiveBench Jan 2025 to Sep 2026
- 18 months o1-mini on AgentIF May 2025 to now
- 18 months GLM-4 on LongBench Jan 2024 to Jul 2025
- 17 months Qwen2-72B on AlGhafa Jun 2024 to Nov 2025
- 16 months Claude 3.5 Sonnet on Aider-Edit Jun 2024 to Oct 2025
- 16 months Gemma 1.1 2B IT on Vectara Hallucination Leaderboard (HHEM) Apr 2024 to Jul 2025
- 15 months o1-preview on KOR-Bench Oct 2024 to Dec 2025
OpenAI closed out 51 of 152 benchmarks
A close-out goes to the model that first took a benchmark past 85%. We count these separately from time at number one.
Google DeepMind 17 Qwen 13 Meta 5 DeepSeek 5 Shanghai AI Lab 4 Moonshot AI 3 ByteDance Seed 3 Other labs 3 xAI 2 
Baidu 2 
StepFun 1 
LG AI Research 1 
Thinking Machines 1 Xiaomi Corp 1 
Meituan 1 MiniMax 1 Physical Intelligence 1 
Perplexity 1 
Microsoft 1 Z.ai 1 By category
Time at number one in each category, split by lab. Click one to see its benchmarks.
OpenAI held number one 36% of the contested time
Each lab’s share of the days its models held number one, across 310 benchmarks with enough scores, independent runs only. We count only contested time. A benchmark’s clock stops when its best score passes 85%, or at its newest score once 45 days go by without another. A new version doesn’t stop the old one’s clock while people still test it.
Google DeepMind 20% Qwen 6% xAI 3% Meta 2% Other labs 2% DeepSeek 1% ByteDance Seed 1% Physical Intelligence 1% Shanghai AI Lab 1% 
Alibaba Group 1% 
Technology Innovation Institute 1% 
Fixie.ai (Ultravox) 1% The longest runs at the top
Independent runs only, counted from January 2023 and only while the benchmark was contested.
- 19 months Gemma 1.1 2B IT on Vectara Hallucination Leaderboard (HHEM) Apr 2024 to Nov 2025
- 17 months Gemini 2.5 Pro on LiveBench Mar 2025 to Sep 2026
- 14 months Qwen2-72B on AlGhafa Jun 2024 to Aug 2025
- 14 months gpt-oss-120b on FORTRESS Aug 2025 to Sep 2026
- 13 months Gemini 1.5 Pro on Video-MME May 2024 to Jun 2025
- 11 months o3 on CharXiv Apr 2025 to Mar 2026
- 11 months Qwen2.5-72B on ArabicMMLU Sep 2024 to Aug 2025
- 11 months Qwen2.5-72B on ArbMMLU-HT Sep 2024 to Aug 2025
OpenAI closed out 39 of 86 benchmarks
A close-out goes to the model that first took a benchmark past 85%. We count these separately from time at number one.
Google DeepMind 12 Qwen 4 Shanghai AI Lab 4 Meta 3 Xiaomi Corp 1 xAI 1 By category
Time at number one in each category, split by lab. Click one to see its benchmarks.