The Pulse

Where AI models stand today

Best score to date on the Epoch Capabilities Index, a composite of many benchmarks. Select a step to see the model that set it.

Closed weights (best to date)Open weights (best to date)
Best Epoch Capabilities Index score to dateStep chart from January 2024 to 22 Nov 2026. Closed-weight best: Claude Fable 5.2, 173.4. Open-weight best: Kimi K3.1 Pro, 161.2. A table of leaders follows the chart.120130140150160170Jan 2024Jul 2024Jan 2025Jul 2025Jan 2026Jul 2026todayClaude Fable 5.2 173.4Kimi K3.1 Pro 161.2
Claude Fable 5.2Anthropic · released 9 Nov 2026ECI 173.4Closed weights
Data: Epoch AI, Capabilities & benchmarking (CC BY 4.0), last fetched 22 Nov 2026. A capability index, not real-world usefulness or adoption. CSV · JSON · · · Instagram size

Month by month

Pick a month to see its leaders, releases and movement. The table shows the frontier score, the best open model and the gap at each month's end. Scores: Epoch Capabilities Index (CC BY 4.0).

Download (CSV)
MONTHFRONTIERMODELSBEST OPENGAPΔ GAPRELEASES
2026-11173.4Claude Fable 5.2Kimi K3.1 Pro (open)161.212.2+4.24
2026-10169.2GPT-6 AstraKimi K3.1 Pro (open)161.28.0-2.610
2026-09169.2GPT-6 AstraKimi K3 (open)158.610.6+4.512
2026-08164.7Claude Opus 5Kimi K3 (open)158.66.10.08
2026-07164.7Claude Opus 5Kimi K3 (open)158.66.1-5.012
2026-06164.2Claude Fable 5GLM-5.2 (open)153.111.1-0.17
2026-05162.8GPT-5.5 ProKimi K2.6 (open)151.611.20.03
2026-04162.8GPT-5.5 ProKimi K2.6 (open)151.611.2-0.316
2026-03160.3GPT-5.4 ProKimi K2.5 (open)148.811.5+2.96
2026-02157.3GPT-5.2 ProKimi K2.5 (open)148.88.60.010
2026-01157.3GPT-5.2 ProKimi K2.5 (open)148.88.6-2.51
2025-12157.3GPT-5.2 ProDeepSeek-V3.2 (open)146.311.1+3.55
2025-11153.0Gemini 3 ProKimi K2 Thinking (open)145.57.6+0.74
2025-10151.3GPT-5 ProDeepSeek-V3.2-Exp (open)144.46.9+1.32
2025-09150.0GPT-5DeepSeek-V3.2-Exp (open)144.45.6-1.25
2025-08150.0GPT-5Qwen3-235B-A22B-Thinking (Jul 2025) (open)143.26.8+2.67
2025-07147.5o3-proQwen3-235B-A22B-Thinking (Jul 2025) (open)143.24.2-2.56
2025-06147.5o3-proDeepSeek-R1 (May 2025) (open)140.76.7+0.45
2025-05147.0o3DeepSeek-R1 (May 2025) (open)140.76.3-2.16
2025-04147.0o3DeepSeek-R1 (open)138.68.4+3.014
2025-03144.0Gemini 2.5 Pro (Mar 2025)DeepSeek-R1 (open)138.65.4+2.94
2025-02141.2o1DeepSeek-R1 (open)138.62.50.04
2025-01141.2o1DeepSeek-R1 (open)138.62.5-7.04
2024-12141.2o1DeepSeek-V3 (open)131.79.5+2.15
2024-11135.3o1-miniLlama 3.1-405B (open)127.97.40.02
2024-10135.3o1-miniLlama 3.1-405B (open)127.97.40.02
2024-09135.3o1-miniLlama 3.1-405B (open)127.97.4+5.312
2024-08130.0Claude 3.5 SonnetLlama 3.1-405B (open)127.92.10.01
2024-07130.0Claude 3.5 SonnetLlama 3.1-405B (open)127.92.1-5.66
2024-06130.0Claude 3.5 SonnetQwen2-72B (open)122.37.7+0.95
2024-05128.1GPT-4o (May 2024)Llama 3-70B (open)121.36.8+2.35
2024-04125.8GPT-4 Turbo (Apr 2024)Llama 3-70B (open)121.34.5-7.98
2024-03125.6Claude 3 OpusMixtral 8x7B (open)113.212.40.02
2024-02125.6Claude 3 OpusMixtral 8x7B (open)113.212.4+0.89
2024-01124.8GPT-4 Turbo (Nov 2023)Mixtral 8x7B (open)113.211.62

Best model by lab

Highest ECI per organisation in Epoch’s data.

LABMODELECI
AnthropicClaude Fable 5.29 Nov 2026 · closed173.4
OpenAIGPT-6.2 Astra17 Nov 2026 · closed172.8
Google DeepMindGemini 4 Carbon10 Nov 2026 · closed169.7
Meta AIMuse Wonder12 Oct 2026 · closed161.7
MoonshotKimi K3.1 Pro23 Oct 2026 · open161.2
DeepSeekDeepSeek V4.1 Flash 101717 Oct 2026 · open159.4
xAIGrok 4.612 Aug 2026 · closed157.8
AlibabaQwen 3.8 Max2 Aug 2026 · closed157.6
Z.ai (Zhipu AI)GLM-5.314 Aug 2026 · open157.1
Mistral AIMistral Large 429 Oct 2026 · open154.6

Source: Epoch AI, Capabilities Index (CC BY 4.0).

Tools builders are discussing

All tools →

Distinct publishers mentioning each tool, 16 Nov 2026–22 Nov 2026. Attention, not adoption.

Claude94Claude Code29ChatGPT20Gemini16Higgsfield7Codex7n8n5DeepSeek4

How to read the Pulse

What the index is

Epoch’s Capabilities Index combines many public benchmarks into one score so models can be compared over time.

What it is not

It does not measure cost, speed, reliability on your task, or how many people use a model. Check the speedruns and our own tests for that.

Freshness

Updated each time the morning run collects data. If a fetch fails we keep the last good copy and show its date. Last successful fetch: 22 Nov 2026.

Release timeline →History & trends →Open models →