AI building AI · 1 Jan 2023 – 22 Nov 2026
The measurements
The headlines say AI is starting to speed up the making of AI. This page shows what the groups that test AI models publish as data, and how fast the best model has been moving, worked out the same way every day. It shows what models can do on tests. It does not show how much of AI research AI does.
18 Points a year the best model's gain on Epoch's index, last 12 months
1.37× Against the year before 13 points a year then
129 Days to double METR's task length, its own fit from 2023
1 Past METR's ceiling model beyond what its tests can measure
Week by week
The questions
The race to build superintelligence AI that improves itself When will AGI arrive? Stopping dangerous superintelligence Testing claims of AGI The business bet on AGI Countries preparing for AGI Chips and cloud for AGI
What is said
The measurements
Epoch Capabilities Index 258 models · Feb 2023 – 17 Nov 2026
One score per model on a single scale that combines many benchmarks, published by Epoch AI . Each dot is a model on its release date; the line joins the models that set a new best on their day.
In the 12 months to 17 November 2026, the best model gained 17.8 points a year (a straight line through 9 record-setting models), against 12.9 points a year in the 12 months before (6 models): 1.37× .
50 100 150
2024 2025 2026
LLaMA-7B (Meta AI) · 24 Feb 2023 · 77.3 (range 63.4–88.9) LLaMA-13B (Meta AI) · 24 Feb 2023 · 85.2 (range 73.4–95.4) LLaMA-33B (Meta AI) · 24 Feb 2023 · 96.5 (range 87.4–104.5) Cerebras-GPT-13B (Cerebras Systems) · 20 Mar 2023 · 48.9 (range 26.3–67.1) Dolly 2.0-12b (Databricks) · 11 Apr 2023 · 55.5 (range 34.5–72.6) vicuna-13b-v1.1 · 12 Apr 2023 · 76.1 (range 60.8–88.6) Falcon-7B (Technology Innovation Institute) · 24 Apr 2023 · 73.1 (range 58.2–85.4) RedPajama-INCITE-7B-Base · 4 May 2023 · 57.7 (range 37.9–74.4) MPT-7B (MosaicML) · 5 May 2023 · 73.4 (range 58.2–86.2) PaLM 2-L · 17 May 2023 · 114.0 (range 98.4–130.5) PaLM 2-M · 17 May 2023 · 97.9 (range 84.6–109.7) PaLM 2-S · 17 May 2023 · 93.8 (range 80.7–105.7) Falcon-40B (Technology Innovation Institute) · 25 May 2023 · 90.7 (range 80.0–100.0) open_llama_7b · 7 Jun 2023 · 64.4 (range 46.0–79.5) GPT-3.5 Turbo (Jun 2023) (OpenAI) · 13 Jun 2023 · 106.0 (range 97.6–113.4) GPT-4 (Jun 2023) (OpenAI) · 13 Jun 2023 · 119.9 (range 116.2–123.3) MPT-30B (MosaicML) · 22 Jun 2023 · 85.2 (range 73.3–95.3) chatglm2-6b · 24 Jun 2023 · 73.7 (range 58.4–86.8) XGen-7B (Salesforce) · 27 Jun 2023 · 70.3 (range 54.3–84.2) internlm-7b · 5 Jul 2023 · 84.6 (range 72.5–94.8) Claude 2 (Anthropic) · 11 Jul 2023 · 117.2 (range 112.8–121.2) Llama 2-70B (Meta AI) · 18 Jul 2023 · 105.7 (range 99.7–111.0) Llama 2-13B (Meta AI) · 18 Jul 2023 · 96.5 (range 87.4–104.2) Llama 2-7B (Meta AI) · 18 Jul 2023 · 83.2 (range 70.6–93.8) Llama 2-34B (Meta AI) · 18 Jul 2023 · 93.9 (range 83.8–102.6) Stable Beluga 2 (Stability AI) · 20 Jul 2023 · 109.4 (range 102.0–116.1) Claude Instant (Anthropic) · 9 Aug 2023 · 114.0 (range 107.0–120.3) Falcon-180B (Technology Innovation Institute) · 6 Sep 2023 · 107.0 (range 98.2–115.5) Baichuan2-13B (Baichuan) · 6 Sep 2023 · 91.8 (range 80.8–100.9) Phi-1.5 (Microsoft) · 11 Sep 2023 · 68.4 (range 50.5–82.9) internlm-20b · 18 Sep 2023 · 98.9 (range 89.1–107.5) Baichuan 2-7B (Baichuan) · 20 Sep 2023 · 83.6 (range 70.5–94.5) Qwen-14B (Alibaba) · 24 Sep 2023 · 101.8 (range 93.2–109.3) Mistral 7B v0.1 (Mistral AI) · 27 Sep 2023 · 102.7 (range 94.9–109.1) Qwen-7B (Alibaba) · 28 Sep 2023 · 89.1 (range 77.9–98.7) DeepSeek Coder 33B (DeepSeek,Peking University) · 2 Nov 2023 · 76.3 (range 60.4–89.5) DeepSeek Coder 6.7B (DeepSeek,Peking University) · 2 Nov 2023 · 67.5 (range 49.2–82.2) GPT-3.5 Turbo (Nov 2023) (OpenAI) · 6 Nov 2023 · 112.7 (range 107.4–117.2) Yi 6B (01.AI) · 22 Nov 2023 · 94.3 (range 84.1–103.2) Qwen-1_8B · 30 Nov 2023 · 59.6 (range 40.2–75.7) Mixtral 8x7B (Mistral AI) · 11 Dec 2023 · 113.2 (range 108.1–117.6) Phi-2 (Microsoft) · 12 Dec 2023 · 96.2 (range 86.9–104.3) GPT-3.5 Turbo (Jan 2024) (OpenAI) · 25 Jan 2024 · 112.9 (range 107.9–117.3) StarCoder 2 7B (Hugging Face,ServiceNow,Nvidia,BigCode) · 20 Feb 2024 · 72.4 (range 54.9–86.6) StarCoder 2 15B (Hugging Face,ServiceNow,Nvidia,BigCode) · 20 Feb 2024 · 96.1 (range 84.0–106.0) Gemma 7B (Google DeepMind) · 21 Feb 2024 · 103.2 (range 95.4–109.7) Gemma 2B (Google DeepMind) · 21 Feb 2024 · 75.7 (range 61.1–87.6) StarCoder 2 3B (Hugging Face,ServiceNow,Nvidia,BigCode) · 22 Feb 2024 · 66.6 (range 47.9–81.7) Mistral Large (Mistral AI) · 26 Feb 2024 · 120.0 (range 115.9–123.6) Nemotron-4 15B (Nvidia) · 26 Feb 2024 · 99.1 (range 90.0–107.0) Claude 3 Sonnet (Anthropic) · 29 Feb 2024 · 119.7 (range 115.5–123.2) Yi-9B · 1 Mar 2024 · 102.2 (range 91.3–111.8) Claude 3 Haiku (Anthropic) · 7 Mar 2024 · 117.8 (range 113.5–121.6) CodeQwen1.5-7B · 15 Apr 2024 · 74.8 (range 57.5–88.6) Mixtral 8x22B (Mistral AI) · 17 Apr 2024 · 118.7 (range 114.5–122.5) Llama 3-70B (Meta AI) · 18 Apr 2024 · 121.3 (range 117.5–124.7) Llama 3-8B (Meta AI) · 18 Apr 2024 · 109.1 (range 103.4–114.0) phi-3-medium 14B (Microsoft) · 23 Apr 2024 · 114.1 (range 109.2–118.5) phi-3-mini 3.8B (Microsoft) · 23 Apr 2024 · 110.5 (range 104.0–116.1) phi-3-small 7.4B (Microsoft) · 23 Apr 2024 · 116.3 (range 110.5–121.9) DeepSeek-V2 (MoE-236B, May 2024) (DeepSeek) · 7 May 2024 · 120.8 (range 114.0–127.6) Falcon 2 11B (Technology Innovation Institute) · 9 May 2024 · 99.2 (range 88.4–109.0) Gemini 1.5 Pro (May 2024) (Google DeepMind) · 14 May 2024 · 125.4 (range 122.1–128.3) Gemini 1.5 Flash (May 2024) (Google DeepMind) · 23 May 2024 · 120.6 (range 116.7–124.1) Qwen2-72B (Alibaba) · 7 Jun 2024 · 122.4 (range 118.6–125.6) DeepSeek-Coder-V2-Lite-Base · 13 Jun 2024 · 99.4 (range 88.8–108.4) Gemma 2 9B (Google DeepMind) · 24 Jun 2024 · 112.4 (range 106.8–117.3) Gemma 2 27B (Google DeepMind) · 24 Jun 2024 · 118.6 (range 114.2–122.4) GPT-4o mini (OpenAI) · 18 Jul 2024 · 123.8 (range 120.5–126.8) Mistral NeMo (Mistral AI) · 18 Jul 2024 · 113.9 (range 108.6–118.4) Llama 3.1-70B (Meta AI) · 23 Jul 2024 · 124.7 (range 121.4–127.7) Llama 3.1-8B (Meta AI) · 23 Jul 2024 · 110.9 (range 105.1–115.9) Llama 3.1-405B (Meta AI) · 23 Jul 2024 · 127.9 (range 125.0–130.5) Mistral Large 2 (Jul 2024) (Mistral AI) · 24 Jul 2024 · 126.4 (range 123.1–129.4) GPT-4o (Aug 2024) (OpenAI) · 6 Aug 2024 · 127.2 (range 124.1–129.9) o1-preview (OpenAI) · 12 Sep 2024 · 133.6 (range 131.1–135.9) Qwen2.5-Coder-0.5B · 18 Sep 2024 · 68.1 (range 48.4–83.9) Qwen2.5-Coder-32B (Alibaba) · 18 Sep 2024 · 118.7 (range 109.5–127.8) Qwen2.5-Coder-14B · 18 Sep 2024 · 113.7 (range 105.2–122.3) Qwen2.5-Coder (1.5B) (Alibaba) · 18 Sep 2024 · 89.1 (range 76.0–100.1) Qwen2.5-Coder-3B · 18 Sep 2024 · 96.8 (range 82.9–109.1) Qwen2.5-Coder (7B) (Alibaba) · 18 Sep 2024 · 107.4 (range 97.7–116.1) Qwen2.5-72B (Alibaba) · 19 Sep 2024 · 127.2 (range 124.4–129.8) Gemini 1.5 Pro (Sept 2024) (Google DeepMind) · 24 Sep 2024 · 130.5 (range 127.8–132.8) Llama 3.2 90B (Meta AI) · 24 Sep 2024 · 122.9 (range 119.2–126.3) Gemini 1.5 Flash (Sep 2024) (Google DeepMind) · 24 Sep 2024 · 127.0 (range 123.9–129.9) Claude 3.5 Sonnet (October 2024) (Anthropic) · 22 Oct 2024 · 133.4 (range 131.2–135.5) Claude 3.5 Haiku (Anthropic) · 22 Oct 2024 · 124.8 (range 121.5–127.7) GPT-4o (Nov 2024) (OpenAI) · 20 Nov 2024 · 128.8 (range 126.1–131.2) INTELLECT-1 (Prime Intellect,Hugging Face,Arcee AI) · 29 Nov 2024 · 83.4 (range 70.8–94.0) Llama 3.3 70B (Meta AI) · 6 Dec 2024 · 126.2 (range 123.0–129.0) Grok-2 (Dec 2024) (xAI) · 12 Dec 2024 · 129.8 (range 127.0–132.4) Phi-4 (Microsoft Research) · 12 Dec 2024 · 129.2 (range 126.1–132.0) DeepSeek-V3 (DeepSeek) · 26 Dec 2024 · 131.7 (range 129.2–134.0) DeepSeek-R1 (DeepSeek) · 20 Jan 2025 · 138.6 (range 136.6–140.5) Gemini 2.0 Flash Thinking (Jan 2025) (Google DeepMind,Google) · 21 Jan 2025 · 133.6 (range 130.7–136.4) Qwen2.5-Max (Alibaba) · 25 Jan 2025 · 131.2 (range 128.4–133.7) o3-mini (OpenAI) · 31 Jan 2025 · 140.6 (range 138.9–142.3) Gemini 2.0 Flash (Feb 2025) (Google DeepMind,Google) · 5 Feb 2025 · 134.0 (range 131.7–136.2) Gemini 2.0 Pro (Google DeepMind) · 5 Feb 2025 · 135.4 (range 132.6–138.0) Claude 3.7 Sonnet (Anthropic) · 24 Feb 2025 · 141.0 (range 139.4–142.6) GPT-4.5 (OpenAI) · 27 Feb 2025 · 136.5 (range 134.4–138.4) QwQ-32B (Alibaba) · 5 Mar 2025 · 136.4 (range 133.7–139.1) Gemma 3 27B (Google DeepMind) · 12 Mar 2025 · 127.6 (range 124.4–130.4) DeepSeek-V3 (Mar 2025) (DeepSeek) · 24 Mar 2025 · 135.3 (range 133.0–137.4) Llama 4 Scout (Meta AI) · 5 Apr 2025 · 127.8 (range 124.9–130.4) Llama 4 Maverick (Meta AI) · 6 Apr 2025 · 132.2 (range 129.9–134.3) Grok 3 (xAI) · 9 Apr 2025 · 137.6 (range 135.6–139.4) GPT-4.1 mini (OpenAI) · 14 Apr 2025 · 134.2 (range 132.1–136.2) GPT-4.1 (OpenAI) · 14 Apr 2025 · 136.1 (range 134.2–138.0) GPT-4.1 nano (OpenAI) · 14 Apr 2025 · 126.8 (range 123.8–129.5) o4-mini (OpenAI) · 16 Apr 2025 · 145.2 (range 143.7–146.7) Gemini 2.5 Flash (Apr 2025) (Google DeepMind) · 17 Apr 2025 · 140.7 (range 138.4–142.9) Qwen3-8B (Alibaba) · 28 Apr 2025 · 132.1 (range 128.9–134.9) Qwen3-235B-A22B (Alibaba) · 28 Apr 2025 · 138.4 (range 136.0–140.7) Qwen3-14B (Alibaba) · 29 Apr 2025 · 136.1 (range 133.3–138.7) Qwen3-32B (Alibaba) · 29 Apr 2025 · 136.9 (range 134.4–139.1) Qwen3-30B-A3B (Alibaba) · 29 Apr 2025 · 134.2 (range 131.3–136.7) Gemini 2.5 Pro (May 2025) (Google DeepMind) · 6 May 2025 · 140.6 (range 138.3–143.0) Mistral Medium 3 (Mistral AI) · 7 May 2025 · 132.8 (range 130.1–135.3) Gemini 2.5 Flash (May 2025) (Google DeepMind) · 20 May 2025 · 141.6 (range 139.4–143.6) Claude Sonnet 4 (Anthropic) · 22 May 2025 · 141.8 (range 140.3–143.3) Claude Opus 4 (Anthropic) · 22 May 2025 · 142.9 (range 141.3–144.4) DeepSeek-R1 (May 2025) (DeepSeek) · 28 May 2025 · 140.7 (range 138.9–142.4) Gemini 2.5 Pro (Jun 2025) (Google DeepMind) · 5 Jun 2025 · 144.9 (range 143.5–146.3) Gemini 2.5 Flash-Lite (Jun 2025) (Google DeepMind) · 17 Jun 2025 · 133.0 (range 129.3–136.7) Gemini 2.5 Flash (Jun 2025) (Google DeepMind) · 17 Jun 2025 · 140.8 (range 138.3–143.1) Grok-3 mini (xAI) · 24 Jun 2025 · 139.9 (range 138.0–141.7) Grok 4 (xAI) · 9 Jul 2025 · 146.9 (range 145.5–148.4) Kimi K2 (Jul 2025) (Moonshot) · 12 Jul 2025 · 141.3 (range 138.5–143.9) Qwen3-235B-A22B-Instruct (Jul 2025) (Alibaba) · 25 Jul 2025 · 138.0 (range 135.7–140.2) Qwen3-235B-A22B-Thinking (Jul 2025) (Alibaba) · 25 Jul 2025 · 143.2 (range 141.3–145.1) Qwen3-30B-A3B-Instruct (Jul 2025) (Alibaba) · 29 Jul 2025 · 134.9 (range 131.9–137.7) Qwen3-30B-A3B-Thinking (Jul 2025) (Alibaba) · 30 Jul 2025 · 138.3 (range 135.8–140.8) Claude Opus 4.1 (Anthropic) · 5 Aug 2025 · 144.8 (range 143.0–146.5) gpt-oss-20b (OpenAI) · 5 Aug 2025 · 134.9 (range 132.3–137.3) gpt-oss-120b (OpenAI) · 5 Aug 2025 · 140.2 (range 138.3–142.0) GPT-5 nano (OpenAI) · 7 Aug 2025 · 139.1 (range 137.1–140.9) GPT-5 mini (OpenAI) · 7 Aug 2025 · 146.0 (range 144.5–147.6) DeepSeek-V3.1 (DeepSeek) · 21 Aug 2025 · 140.9 (range 137.7–144.0) Grok 4 Fast (xAI) · 19 Sep 2025 · 144.3 (range 142.2–146.4) Qwen3-Max (Alibaba) · 24 Sep 2025 · 142.4 (range 140.4–144.5) Gemini 2.5 Flash (Sep 2025) (Google DeepMind) · 25 Sep 2025 · 141.9 (range 139.2–144.5) Claude Sonnet 4.5 (Anthropic) · 29 Sep 2025 · 147.6 (range 146.1–149.1) DeepSeek-V3.2-Exp (DeepSeek) · 29 Sep 2025 · 144.4 (range 142.2–146.7) Claude Haiku 4.5 (Anthropic) · 15 Oct 2025 · 142.7 (range 141.0–144.4) Kimi K2 Thinking (Moonshot) · 6 Nov 2025 · 145.5 (range 143.7–147.3) GPT-5.1 (OpenAI) · 13 Nov 2025 · 150.9 (range 149.4–152.6) Claude Opus 4.5 (Anthropic) · 24 Nov 2025 · 151.9 (range 150.4–153.5) DeepSeek-V3.2 (DeepSeek) · 1 Dec 2025 · 146.3 (range 144.7–147.8) GPT-5.2 (OpenAI) · 11 Dec 2025 · 154.5 (range 152.9–156.3) Gemini 3 Flash (Google DeepMind) · 17 Dec 2025 · 152.4 (range 150.8–154.1) GLM-4.7 (Z.ai (Zhipu AI)) · 22 Dec 2025 · 143.3 (range 141.4–145.3) Kimi K2.5 (Moonshot) · 27 Jan 2026 · 148.8 (range 147.3–150.3) Claude Opus 4.6 (Anthropic) · 5 Feb 2026 · 156.3 (range 154.6–158.3) GLM-5 (Z.ai (Zhipu AI)) · 11 Feb 2026 · 146.5 (range 144.9–148.1) Qwen3.5 397B-A17B (Alibaba) · 13 Feb 2026 · 148.0 (range 145.9–150.1) Qwen 3.5 Plus (hosted 397B-A17B) (Alibaba) · 16 Feb 2026 · 147.3 (range 145.4–149.2) Grok 4.20 (xAI) · 17 Feb 2026 · 152.3 (range 150.5–154.1) Claude Sonnet 4.6 (Anthropic) · 17 Feb 2026 · 152.9 (range 151.2–154.5) Gemini 3.1 Pro (Google DeepMind) · 19 Feb 2026 · 156.3 (range 154.5–158.2) Qwen3.5-9B (Alibaba) · 24 Feb 2026 · 139.5 (range 137.3–141.7) Qwen3.5-35B-A3B (Alibaba) · 24 Feb 2026 · 144.1 (range 141.5–146.6) Qwen 3.5 Flash (hosted 35B-A3B) (Alibaba) · 25 Feb 2026 · 143.8 (range 141.9–145.7) Gemini 3.1 Flash-Lite (Google) · 3 Mar 2026 · 145.0 (range 142.8–147.1) GPT-5.4 (OpenAI) · 5 Mar 2026 · 157.9 (range 156.0–160.0) GPT-5.4 Nano (OpenAI) · 17 Mar 2026 · 146.4 (range 144.7–148.0) GPT-5.4 Mini (OpenAI) · 17 Mar 2026 · 149.4 (range 147.7–151.1) Qwen 3.6 Plus (Alibaba) · 31 Mar 2026 · 148.9 (range 146.9–150.9) Gemma 4 31B IT (Google DeepMind) · 2 Apr 2026 · 144.1 (range 141.7–146.6) Gemma 4 26B A4B (Google DeepMind) · 2 Apr 2026 · 141.6 (range 139.1–143.9) GLM-5.1 (Z.ai (Zhipu AI)) · 7 Apr 2026 · 150.7 (range 148.7–152.8) Muse Spark (Meta AI) · 8 Apr 2026 · 153.2 (range 151.0–155.5) Qwen 3.6 35B-A3B (Alibaba) · 14 Apr 2026 · 144.1 (range 142.2–145.9) Claude Opus 4.7 (Anthropic) · 16 Apr 2026 · 157.5 (range 155.6–159.6) Grok 4.3 Beta (xAI) · 17 Apr 2026 · 149.9 (range 148.0–151.9) Kimi K2.6 (Moonshot) · 20 Apr 2026 · 151.6 (range 149.9–153.4) Qwen 3.6 Max (Preview) (Alibaba) · 20 Apr 2026 · 149.9 (range 147.9–151.9) Qwen3.6 27B (Alibaba) · 22 Apr 2026 · 146.7 (range 144.6–148.9) GPT-5.5 (OpenAI) · 23 Apr 2026 · 160.2 (range 158.2–162.4) DeepSeek-V4-Flash (DeepSeek) · 24 Apr 2026 · 146.9 (range 143.7–150.2) DeepSeek-V4-Pro (DeepSeek) · 24 Apr 2026 · 150.2 (range 148.4–152.0) Qwen 3.6 Flash (Alibaba) · 27 Apr 2026 · 143.7 (range 141.8–145.6) Mistral Medium 3.5 (Mistral AI) · 28 Apr 2026 · 141.6 (range 138.3–144.7) Gemini 3.5 Flash (Google DeepMind) · 19 May 2026 · 155.3 (range 153.5–157.3) Qwen3.7-Max (Alibaba) · 19 May 2026 · 154.6 (range 152.6–156.8) Claude Opus 4.8 (Anthropic) · 28 May 2026 · 159.9 (range 157.8–162.2) MiniMax-M3 (MiniMax) · 1 Jun 2026 · 147.8 (range 145.7–150.0) Qwen3.7-Plus (Alibaba) · 2 Jun 2026 · 147.9 (range 146.0–150.0) Nemotron 3 Ultra (Nvidia) · 4 Jun 2026 · 147.0 (range 144.9–149.2) Kimi K2.7 Code (Moonshot) · 12 Jun 2026 · 151.2 (range 149.2–153.3) GLM-5.2 (Z.ai (Zhipu AI)) · 16 Jun 2026 · 153.1 (range 151.3–154.9) Claude Sonnet 5 (Anthropic) · 30 Jun 2026 · 156.9 (range 154.9–159.1) Grok 4.5 (xAI) · 8 Jul 2026 · 155.0 (range 153.1–156.9) Muse Spark 1.1 (Meta AI) · 9 Jul 2026 · 154.9 (range 152.4–157.7) GPT-5.6 Terra (OpenAI) · 9 Jul 2026 · 160.8 (range 158.5–163.3) GPT-5.6 Sol (OpenAI) · 9 Jul 2026 · 163.0 (range 160.7–165.6) GPT-5.6 Luna (OpenAI) · 9 Jul 2026 · 156.7 (range 154.9–158.8) Inkling (Thinking Machines) · 15 Jul 2026 · 148.4 (range 146.8–150.1) Inkling-Small (Thinking Machines) · 15 Jul 2026 · 150.8 (range 149.0–152.5) Kimi K3 (Moonshot) · 16 Jul 2026 · 158.6 (range 156.7–160.8) Gemini 3.6 Flash (Google DeepMind) · 21 Jul 2026 · 155.2 (range 153.4–157.3) Gemini 3.5 Flash-Lite (Google DeepMind) · 21 Jul 2026 · 145.6 (range 144.0–147.2) DeepSeek V4 Flash 0731 (DeepSeek) · 31 Jul 2026 · 154.4 (range 152.6–156.4) Qwen 3.8 Max (Alibaba) · 2 Aug 2026 · 157.6 (range 155.5–159.9) Muse Spark 1.2 (Meta AI) · 5 Aug 2026 · 155.0 (range 152.7–157.5) Grok 4.6 (xAI) · 12 Aug 2026 · 157.8 (range 155.9–160.0) DeepSeek V4 Pro 0813 (DeepSeek) · 13 Aug 2026 · 156.0 (range 154.1–158.1) Gemini 3.7 Flash (Google DeepMind) · 13 Aug 2026 · 158.4 (range 156.5–160.6) Qwen 3.8 27B (Alibaba) · 14 Aug 2026 · 152.6 (range 150.7–154.7) GLM-5.3 (Z.ai (Zhipu AI)) · 14 Aug 2026 · 157.1 (range 155.0–159.4) GLM-5.3-Flash (Z.ai (Zhipu AI)) · 20 Aug 2026 · 153.8 (range 152.0–155.8) Qwen3.8 Max (0902) (Alibaba) · 1 Sep 2026 · 156.0 (range 153.9–158.3) Gemini 3.8 Flash (Google DeepMind) · 2 Sep 2026 · 158.8 (range 156.8–161.0) Muse Spark 1.3 (Meta AI) · 2 Sep 2026 · 158.2 (range 156.2–160.5) DeepSeek V4.1 Flash (DeepSeek) · 9 Sep 2026 · 157.0 (range 155.1–159.2) Grok 4.7 (xAI) · 21 Sep 2026 · 156.4 (range 154.5–158.5) Claude Opus 5.5 (Anthropic) · 22 Sep 2026 · 168.7 (range 165.9–172.0) GPT-6 Sol (OpenAI) · 22 Sep 2026 · 163.8 (range 161.4–166.6) GPT-6 Luna (OpenAI) · 22 Sep 2026 · 156.5 (range 154.6–158.5) Claude Sonnet 5.5 (Anthropic) · 28 Sep 2026 · 166.6 (range 163.9–169.6) GPT-6.1 Sol (OpenAI) · 29 Sep 2026 · 168.6 (range 165.7–171.9) Beam 1.0 (Reflection) · 5 Oct 2026 · 150.8 (range 148.6–153.1) Mistral Large 4 Preview (Mistral AI) · 6 Oct 2026 · 152.7 (range 150.8–154.7) Claude Haiku 5.5 (Anthropic) · 7 Oct 2026 · 160.3 (range 158.2–162.7) Muse Wonder (Meta AI) · 12 Oct 2026 · 161.7 (range 159.4–164.3) Gemini 4 Argon (Google DeepMind) · 13 Oct 2026 · 166.5 (range 163.9–169.5) Muse Spark Open (Meta AI) · 15 Oct 2026 · 155.5 (range 153.3–157.8) DeepSeek V4.1 Flash 1017 (DeepSeek) · 17 Oct 2026 · 159.4 (range 157.1–161.9) Kimi K3.1 Pro (Moonshot) · 23 Oct 2026 · 161.2 (range 159.0–163.7) GPT-6.1 Terra (OpenAI) · 26 Oct 2026 · 167.7 (range 164.9–171.0) Mistral Large 4 (Mistral AI) · 29 Oct 2026 · 154.6 (range 151.8–157.4) Gemini 4 Carbon (Google DeepMind) · 10 Nov 2026 · 169.7 (range 166.6–173.1) Gemini 4 Hydrogen (Google DeepMind) · 12 Nov 2026 · 166.8 (range 164.0–170.1) GPT-6.2 Astra (OpenAI) · 17 Nov 2026 · 172.8 (range 169.4–176.9)
LLaMA-65B (Meta AI) · 24 Feb 2023 · 101.5 (range 93.2–108.9) · a new best GPT-4 (Mar 2023) (OpenAI) · 14 Mar 2023 · 121.6 (range 117.6–125.2) · a new best GPT-4 Turbo (Nov 2023) (OpenAI) · 25 Jan 2024 · 124.8 (range 121.5–127.7) · a new best Claude 3 Opus (Anthropic) · 29 Feb 2024 · 125.6 (range 122.5–128.3) · a new best GPT-4 Turbo (Apr 2024) (OpenAI) · 9 Apr 2024 · 125.8 (range 122.5–128.8) · a new best GPT-4o (May 2024) (OpenAI) · 13 May 2024 · 128.1 (range 125.4–130.7) · a new best Claude 3.5 Sonnet (Anthropic) · 20 Jun 2024 · 130.0 · a new best o1-mini (OpenAI) · 12 Sep 2024 · 135.3 (range 133.0–137.3) · a new best o1 (OpenAI) · 17 Dec 2024 · 141.2 (range 139.5–142.8) · a new best Gemini 2.5 Pro (Mar 2025) (Google DeepMind) · 31 Mar 2025 · 144.0 (range 142.0–146.0) · a new best o3 (OpenAI) · 16 Apr 2025 · 147.0 (range 145.5–148.6) · a new best o3-pro (OpenAI) · 10 Jun 2025 · 147.5 (range 145.7–149.4) · a new best GPT-5 (OpenAI) · 7 Aug 2025 · 150.0 · a new best GPT-5 Pro (OpenAI) · 7 Oct 2025 · 151.3 (range 149.4–153.4) · a new best Gemini 3 Pro (Google DeepMind) · 18 Nov 2025 · 153.0 (range 151.5–154.8) · a new best GPT-5.2 Pro (OpenAI) · 11 Dec 2025 · 157.4 (range 155.1–160.0) · a new best GPT-5.4 Pro (OpenAI) · 5 Mar 2026 · 160.3 (range 157.9–163.0) · a new best GPT-5.5 Pro (OpenAI) · 23 Apr 2026 · 162.8 (range 160.1–165.9) · a new best Claude Fable 5 (Anthropic) · 9 Jun 2026 · 164.2 (range 161.8–166.9) · a new best Claude Opus 5 (Anthropic) · 24 Jul 2026 · 164.7 (range 162.3–167.5) · a new best Claude Fable 5.1 (Anthropic) · 1 Sep 2026 · 167.1 (range 164.4–170.1) · a new best GPT-6 Astra (OpenAI) · 3 Sep 2026 · 169.2 (range 166.4–172.5) · a new best Claude Fable 5.2 (Anthropic) · 9 Nov 2026 · 173.4 (range 170.0–177.2) · a new best
Claude Fable 5.2
The pace, 12 months at a time
12 months to 31 Dec 2024: 19.0 points a year, 7 models 19 Dec 24 12 months to 31 Mar 2025: 19.3 points a year, 6 models 19 Mar 25 12 months to 30 Jun 2025: 16.5 points a year, 5 models 17 Jun 25 12 months to 30 Sep 2025: 13.9 points a year, 5 models 14 Sep 25 12 months to 31 Dec 2025: 15.0 points a year, 7 models 15 Dec 25 12 months to 31 Mar 2026: 15.6 points a year, 7 models 16 Mar 26 12 months to 30 Jun 2026: 18.4 points a year, 7 models 18 Jun 26 12 months to 30 Sep 2026: 17.4 points a year, 9 models 17 Sep 26 12 months to 17 Nov 2026: 17.8 points a year, 9 models 18 Nov 26
Each bar is the 12 months ending on its date: points a year, fitted through the models that set a new best in that window. A dash: fewer than three such models. Data: Epoch's file , fetched 2026-11-22.
Training compute 19 models · on the index's dates
The computing used to train each model, in operations (FLOP), from Epoch AI's list of frontier models . Drawn on the same dates as the index above so the two line up; the scale goes up by multiples of ten. Many labs do not publish this, so Epoch's figures for some models are estimates and newer models are often missing.
Too few record-setting models in the last 12 months to fit a pace.
10²⁵ 10²⁶ 10²⁷
2024 2025 2026
SenseChat (SenseTime) · 10 Apr 2023 · 3.9×10²⁴ FLOP PaLM 2 (Google) · 10 May 2023 · 7.3×10²⁴ FLOP GPT-4 (Jun 2023) (OpenAI) · 13 Jun 2023 · 2.1×10²⁵ FLOP Claude 2 (Anthropic) · 11 Jul 2023 · 3.9×10²⁴ FLOP Falcon-180B (Technology Innovation Institute) · 6 Sep 2023 · 3.8×10²⁴ FLOP Amazon Titan (Amazon) · 28 Sep 2023 · 4.8×10²⁴ FLOP Inflection-2 (Inflection AI) · 22 Nov 2023 · 1.0×10²⁵ FLOP Nemotron-4 340B (Nvidia) · 14 Jun 2024 · 1.8×10²⁵ FLOP Claude 3.5 Sonnet (Anthropic) · 20 Jun 2024 · 2.7×10²⁵ FLOP Llama 3.1-405B (Meta AI) · 23 Jul 2024 · 3.8×10²⁵ FLOP Grok-2 (xAI) · 13 Aug 2024 · 3.0×10²⁵ FLOP Llama Nemotron Ultra 253B (Nvidia) · 18 Mar 2025 · 3.9×10²⁵ FLOP Llama 4 Behemoth (preview) (Meta AI) · 5 Apr 2025 · 5.2×10²⁵ FLOP
GPT-4 (Mar 2023) (OpenAI) · 15 Mar 2023 · 2.1×10²⁵ FLOP · a new best Gemini 1.0 Ultra (Google DeepMind) · 6 Dec 2023 · 5.0×10²⁵ FLOP · a new best Grok 3 (xAI) · 17 Feb 2025 · 3.5×10²⁶ FLOP · a new best GPT-4.5 (OpenAI) · 27 Feb 2025 · 3.8×10²⁶ FLOP · a new best Grok 4 (xAI) · 9 Jul 2025 · 5.0×10²⁶ FLOP · a new best GPT-6 Astra (OpenAI) · 3 Sep 2026 · 9.4×10²⁶ FLOP · a new best
GPT-6 Astra
Data: Epoch's file , fetched 2026-11-22.
METR task length 26 models · Feb 2019 – 7 Apr 2026
The length of software task, in a skilled person's time, that a model finishes half the time, measured by METR . As METR draws it, the scale goes up by multiples: each line up is several times the one below; switch to a straight scale to see the same models drawn to size. METR stopped updating these in September 2026, and says lengths above 16 hours can't be measured reliably with its tasks.
From 2023, METR's own fit has the length doubling every 129 days (95% range 104–158). Claude Mythos Preview Early, released 7 April 2026, measured 1,045 minutes: past the 16 hours its tasks can measure, so no 12-month pace can be fitted after it.
Days to double, 12 months at a time shorter is faster
12 months to 30 Jun 2024: doubling every 172 days, 3 models 172 Jun 24 12 months to 30 Sep 2024: doubling every 138 days, 4 models 138 Sep 24 12 months to 31 Dec 2024: doubling every 93 days, 5 models 93 Dec 24 12 months to 31 Mar 2025: doubling every 95 days, 6 models 95 Mar 25 12 months to 30 Jun 2025: doubling every 83 days, 5 models 83 Jun 25 12 months to 30 Sep 2025: doubling every 89 days, 5 models 89 Sep 25 12 months to 31 Dec 2025: doubling every 139 days, 6 models 139 Dec 25 12 months to 31 Mar 2026: doubling every 135 days, 6 models 135 Mar 26 past ceiling Apr 26
Each bar is the 12 months ending on its date, fitted through the models that set a new best in that window. "Past ceiling": the window holds a model METR can't measure, so its top is missing and a fit would read as a slowdown that isn't there. Data: METR's file , fetched 2026-11-22.
Coding and research tests
Four tests closest to the work of building AI, as Epoch AI collects them: fixing real software bugs (SWE-bench Verified), getting jobs done at a command line (Terminal-Bench), long software-engineering projects (FrontierSWE), and making code run faster (AlgoTune, as a speedup over the reference code). A model run several ways keeps its best score. Each covers only the months it has been run.
30 40 50 60 70 80 90
2025 2026
GPT-4.1 (OpenAI) · 14 Apr 2025 · 48.5% gemini-2.5-pro (Google DeepMind) · 17 Jun 2025 · 57.6% GPT-5 mini (medium) (OpenAI) · 7 Aug 2025 · 64.7% GPT-5 (medium) (OpenAI) · 7 Aug 2025 · 71.5% Claude Sonnet 4.5 (no thinking) (Anthropic) · 29 Sep 2025 · 71.3% GPT-5.1 (high) (OpenAI) · 13 Nov 2025 · 68.0% Gemini 3 Pro Preview (Google DeepMind) · 18 Nov 2025 · 72.9% GPT-5.2 (high) (OpenAI) · 11 Dec 2025 · 73.8% Gemini 3 Flash Preview (Google DeepMind) · 17 Dec 2025 · 75.4% Kimi K2.5 (Moonshot) · 27 Jan 2026 · 73.8% GPT-5.3 Codex (high) (OpenAI) · 5 Feb 2026 · 74.8% GLM-5 (Z.ai (Zhipu AI)) · 11 Feb 2026 · 72.1% Claude Sonnet 4.6 (no thinking) (Anthropic) · 17 Feb 2026 · 75.2% Gemini 3.1 Pro Preview (Google DeepMind) · 19 Feb 2026 · 75.6% GPT-5.4 (high) (OpenAI) · 5 Mar 2026 · 76.9% Qwen 3.6 Plus (2026-04-02) (Alibaba) · 31 Mar 2026 · 57.9% GLM-5.1 (Z.ai (Zhipu AI)) · 7 Apr 2026 · 74.2% Qwen 3.6 Max (Preview) (Alibaba) · 20 Apr 2026 · 76.7% Kimi K2.6 (Moonshot) · 20 Apr 2026 · 76.7% GPT-5.5 (xhigh) (OpenAI) · 23 Apr 2026 · 80.6% DeepSeek v4 Pro (max) (DeepSeek) · 24 Apr 2026 · 77.6% Gemini 3.5 Flash (Google DeepMind) · 19 May 2026 · 79.3% Qwen3.7 Max (Alibaba) · 19 May 2026 · 77.3% GLM-5.2 (Z.ai (Zhipu AI)) · 16 Jun 2026 · 78.7%
GPT-4o (Nov 2024) (OpenAI) · 20 Nov 2024 · 31.0% · a new best claude-3-7-sonnet-20250219 (Anthropic) · 24 Feb 2025 · 61.0% · a new best o3 (medium) (OpenAI) · 16 Apr 2025 · 62.3% · a new best Claude Opus 4 (Anthropic) · 22 May 2025 · 70.7% · a new best claude-opus-4-1-20250805 (Anthropic) · 5 Aug 2025 · 73.3% · a new best GPT-5 (high) (OpenAI) · 7 Aug 2025 · 73.6% · a new best Claude Opus 4.5 (no thinking) (Anthropic) · 24 Nov 2025 · 76.7% · a new best Claude Opus 4.6 (no thinking) (Anthropic) · 5 Feb 2026 · 78.7% · a new best Claude Opus 4.7 (max) (Anthropic) · 16 Apr 2026 · 83.5% · a new best
Claude Opus 4.7 (max)
0 20 40 60 80
2026
gemini-2.5-pro (Google DeepMind) · 17 Jun 2025 · 32.6% Gemini 2.5 Flash (Google DeepMind) · 17 Jun 2025 · 17.1% Grok 4 (xAI) · 9 Jul 2025 · 27.2% grok-4-0709 (xAI) · 9 Jul 2025 · 27.2% Kimi K2 Instruct (Moonshot) · 12 Jul 2025 · 27.8% Qwen 3 Coder 480B (Alibaba) · 31 Jul 2025 · 27.2% Qwen3-Coder-480B-A35B-Instruct (Alibaba) · 31 Jul 2025 · 25.4% gpt-oss-120b (unknown thinking) (OpenAI) · 5 Aug 2025 · 18.7% Claude Opus 4.1 (unknown thinking) (Anthropic) · 5 Aug 2025 · 38.0% gpt-oss-20b (OpenAI) · 5 Aug 2025 · 3.4% gpt-oss-120b (OpenAI) · 5 Aug 2025 · 18.7% GPT-5 mini (unknown thinking) (OpenAI) · 7 Aug 2025 · 34.8% GPT-5 nano (unknown thinking) (OpenAI) · 7 Aug 2025 · 21.8% GPT-5 mini (medium) (OpenAI) · 7 Aug 2025 · 31.9% GPT-5 nano (medium) (OpenAI) · 7 Aug 2025 · 11.5% GPT-5 (medium) (OpenAI) · 7 Aug 2025 · 49.6% grok-code-fast-1 · 28 Aug 2025 · 25.8% Grok Code Fast 1 · 28 Aug 2025 · 25.8% GPT-5-codex (OpenAI) · 15 Sep 2025 · 44.3% gemini-2.5-flash-preview-09-2025 (Google DeepMind) · 25 Sep 2025 · 17.1% Claude Sonnet 4.5 (unknown thinking) (Anthropic) · 29 Sep 2025 · 46.5% Claude Sonnet 4.5 (no thinking) (Anthropic) · 29 Sep 2025 · 42.8% GLM-4.6 (Z.ai (Zhipu AI),Tsinghua University) · 30 Sep 2025 · 24.5% claude-haiku-4-5-20251001 (Anthropic) · 15 Oct 2025 · 29.8% Claude Haiku 4.5 (unknown thinking) (Anthropic) · 15 Oct 2025 · 35.5% MiniMax M2 (MiniMax) · 27 Oct 2025 · 30.0% Kimi K2 Thinking (Moonshot) · 6 Nov 2025 · 35.7% GPT-5.1 Codex (OpenAI) · 12 Nov 2025 · 57.8% GPT-5.1 (unknown thinking) (OpenAI) · 13 Nov 2025 · 47.6% GPT-5.1 (medium) (OpenAI) · 13 Nov 2025 · 47.6% GPT-5.1-Codex-Max (OpenAI) · 19 Nov 2025 · 60.4% Claude Opus 4.5 (128k thinking) (Anthropic) · 24 Nov 2025 · 59.1% Claude Opus 4.5 (unknown thinking) (Anthropic) · 24 Nov 2025 · 63.1% Claude Opus 4.5 (no thinking) (Anthropic) · 24 Nov 2025 · 63.1% DeepSeek-V3.2 (Thinking; Novita) (DeepSeek) · 1 Dec 2025 · 39.6% GPT-5.2 (unknown thinking) (OpenAI) · 11 Dec 2025 · 64.9% GPT-5.2 (medium) (OpenAI) · 11 Dec 2025 · 64.9% Gemini 3 Flash Preview (Google DeepMind) · 17 Dec 2025 · 64.3% GPT-5.2 Codex (OpenAI) · 18 Dec 2025 · 66.5% GLM-4.7 (Z.ai (Zhipu AI)) · 22 Dec 2025 · 33.4% MiniMax M2.1 (MiniMax) · 23 Dec 2025 · 36.6% Kimi K2.5 (Moonshot) · 27 Jan 2026 · 43.2% GPT-5.3 Codex (OpenAI) · 5 Feb 2026 · 78.4% Claude Opus 4.6 (no thinking) (Anthropic) · 5 Feb 2026 · 69.9% GLM-5 (Z.ai (Zhipu AI)) · 11 Feb 2026 · 52.4% minimax-m2.5 (MiniMax) · 12 Feb 2026 · 42.7% Minimax m2.5 (MiniMax) · 12 Feb 2026 · 42.2% Claude Sonnet 4.6 (unknown thinking) (Anthropic) · 17 Feb 2026 · 53.4% Grok 4.20 Reasoning (xAI) · 17 Feb 2026 · 57.3% Claude Sonnet 4.6 (no thinking) (Anthropic) · 17 Feb 2026 · 53.4% Qwen3.5 9B (Alibaba) · 24 Feb 2026 · 9.2% MiniMax M2.7 (MiniMax) · 18 Mar 2026 · 45.1% MiniMax M2.7 Highspeed (MiniMax) · 18 Mar 2026 · 42.9% Qwen3.6 35B-A3B (Alibaba) · 14 Apr 2026 · 24.6% Claude Opus 4.7 (unknown) (Anthropic) · 16 Apr 2026 · 80.2%
Gemini 2.5 Pro (Google DeepMind) · 17 Jun 2025 · 32.6% · a new best claude-opus-4-1-20250805 (Anthropic) · 5 Aug 2025 · 38.0% · a new best GPT-5 (unknown thinking) (OpenAI) · 7 Aug 2025 · 49.6% · a new best GPT-5.1 Codex Mini (OpenAI) · 12 Nov 2025 · 61.6% · a new best Gemini 3 Pro Preview (Google DeepMind) · 18 Nov 2025 · 69.4% · a new best Claude Opus 4.6 (unknown thinking) (Anthropic) · 5 Feb 2026 · 79.8% · a new best Gemini 3.1 Pro Preview (Google DeepMind) · 19 Feb 2026 · 80.2% · a new best GPT-5.4 (unknown thinking) (OpenAI) · 5 Mar 2026 · 81.8% · a new best GPT-5.5 (unknown thinking) (OpenAI) · 23 Apr 2026 · 84.7% · a new best
GPT-5.5 (unknown thinking)
FrontierSWE 25 models · Jun 2026 – Nov 2026
0 20 40 60
GPT-5.6 Sol (max) (OpenAI) · 9 Jul 2026 · 32.2% Inkling (xhigh) (Thinking Machines) · 15 Jul 2026 · 4.1% Kimi K3 (max) (Moonshot) · 16 Jul 2026 · 25.9% Qwen3.8 Max (xhigh) (Alibaba) · 2 Aug 2026 · 15.8% Muse Spark 1.2 (xhigh) (Meta AI) · 5 Aug 2026 · 12.0% Grok 4.6 (xhigh) (xAI) · 12 Aug 2026 · 25.3% Gemini 3.7 Flash (high) (Google DeepMind) · 13 Aug 2026 · 20.3% GLM-5.3 (max) (Z.ai (Zhipu AI)) · 14 Aug 2026 · 30.2% GLM-5.3-Flash (max) (Z.ai (Zhipu AI)) · 20 Aug 2026 · 18.1% Qwen3.8 Max (0902) (xhigh) (Alibaba) · 1 Sep 2026 · 17.8% Muse Spark 1.3 (max) (Meta AI) · 2 Sep 2026 · 25.8% Gemini 3.8 Flash (high) (Google DeepMind) · 2 Sep 2026 · 19.6% Grok 4.7 (xhigh) (xAI) · 21 Sep 2026 · 29.5% Claude Opus 5.5 (max) (Anthropic) · 22 Sep 2026 · 62.3% Claude Sonnet 5.5 (max) (Anthropic) · 28 Sep 2026 · 61.9% Claude Haiku 5.5 (max) (Anthropic) · 7 Oct 2026 · 43.8% Gemini 4 Argon (high) (Google DeepMind) · 13 Oct 2026 · 55.0% Inkling 2 Small (xhigh) (Thinking Machines) · 5 Nov 2026 · 25.1% Gemini 4 Carbon (high) (Google DeepMind) · 10 Nov 2026 · 61.2% Gemini 4 Hydrogen (high) (Google DeepMind) · 12 Nov 2026 · 53.1%
Claude Fable 5 (max) (Anthropic) · 9 Jun 2026 · 47.0% · a new best Claude Opus 5 (max) (Anthropic) · 24 Jul 2026 · 52.0% · a new best Claude Fable 5.1 (max) (Anthropic) · 1 Sep 2026 · 56.3% · a new best GPT-6 Astra (max) (OpenAI) · 3 Sep 2026 · 65.5% · a new best Claude Fable 5.2 (max) (Anthropic) · 9 Nov 2026 · 69.9% · a new best
Claude Fable 5.2 (max)
AlgoTune 18 models · Jan 2025 – Mar 2026
0 0.5 1 1.5 2
2026
Claude Opus 4 (Anthropic) · 22 May 2025 · 1.33× Gemini 2.5 Pro (high) (Google DeepMind) · 17 Jun 2025 · 1.51× Qwen3 Coder (high) (Alibaba) · 31 Jul 2025 · 1.44× gpt-oss-120b (high) (OpenAI) · 5 Aug 2025 · 1.41× Claude Opus 4.1 (unknown thinking) (Anthropic) · 5 Aug 2025 · 1.34× GLM-4.5 Thinking (Z.ai (Zhipu AI),Tsinghua University) · 5 Aug 2025 · 1.52× GPT-5 mini (high) (OpenAI) · 7 Aug 2025 · 1.38× GPT-5 (high) (OpenAI) · 7 Aug 2025 · 1.67× Claude Sonnet 4.5 (unknown thinking) (Anthropic) · 29 Sep 2025 · 1.52× GPT-5 Pro (OpenAI) · 7 Oct 2025 · 1.31× Claude Opus 4.5 (unknown thinking) (Anthropic) · 24 Nov 2025 · 1.77× Claude Opus 4.6 (high) (Anthropic) · 5 Feb 2026 · 1.47× Gemini 3.1 Pro Preview (Google DeepMind) · 19 Feb 2026 · 2.02× GPT-5.4 (high) (OpenAI) · 5 Mar 2026 · 1.85×
DeepSeek R1 (high) (DeepSeek) · 20 Jan 2025 · 1.70× · a new best o4-mini (high) (OpenAI) · 16 Apr 2025 · 1.72× · a new best Gemini 3 Pro Preview (Google DeepMind) · 18 Nov 2025 · 1.83× · a new best GPT-5.2 (medium) (OpenAI) · 11 Dec 2025 · 2.05× · a new best
GPT-5.2 (medium)
Data: Epoch's benchmark file .
Faster or slower? Every measure side by side
For each measure: how fast the best model moved in the last 12 months of its data, against the 12 months before, worked out the same way (a straight line through the models that set a new best). Above 1× is faster. A dash means too few models to say.
Measure Last 12 months 12 months before Change Latest model Note
Epoch Capabilities Index
+17.8 points a year (9)
+12.9 points a year (6)
1.37×
17 Nov 2026
METR task length (50% success)
—
doubling every 95 days (6)
—
7 Apr 2026
past METR's ceiling: no pace can be fitted after it; no new model since 7 Apr 2026
Training compute
—
—
—
3 Sep 2026
fewer than 3 new bests in the last 12 months
SWE-bench Verified
+13.4 points a year (5)
+76.9 points a year (4)
0.17×
16 Jun 2026
Terminal-Bench 2.0
+64.7 points a year (9)
—
—
23 Apr 2026
fewer than 3 new bests in the 12 months before; no new model since 23 Apr 2026
FrontierSWE
—
—
—
12 Nov 2026
new bests in the last 12 months span under six months
AlgoTune
doubling every 1248 days (3)
—
—
5 Mar 2026
fewer than 3 new bests in the 12 months before; no new model since 5 Mar 2026
In brackets: how many record-setting models the line rests on. A percentage near 100 has little room left to rise, so its pace slows whatever the models do. Measures stop when their publisher stops running them, so their last 12 months can end long ago.
What this shows, and what it doesn't
These are scores on tests. A model that scores higher is not shown here to have done any AI research. Whether AI is doing more of the work of building AI is a different question; the headlines on the other pages follow what is reported about it.
The comparison of one year's pace with the year before is the one METR's November 2026 assessment of Anthropic made, on Anthropic's own version of this index and with evidence that is not public. This page uses the public models of every lab, so the two numbers measure different things.
Nothing here is projected forward. A pace is a straight line through the models of one 12-month window, with the number of models it rests on.