Model version,Time horizon,average_score,Release date,Organization,Country,Training compute (FLOP),Training compute notes,CI_high,CI_low,Source,Source link,Notes,Time Horizon (80%),METR version,id,date,sota
gpt-4-0314,3.987428,0.293044,2023-03-14,OpenAI,United States of America,2.1e+25,"90% CI: 8.2E+24 to 4.4E+25

NOTE: this is a rough estimate based on public information, much less information than most other systems in the database.

Calculation and confidence intervals here: https://colab.research.google.com/drive/1O99z9b1I5O66bT78r9ScslE_nOj5irN9?usp=sharing",7.906991,1.959555,,,,0.889561,METR-Horizon-v1.1,recGv2XrraPE8opN2,2023-03-14,3.987428
gpt-4-1106-preview,4.044959,0.28905,2023-11-06,OpenAI,United States of America,,Training compute estimated to be 2.2e25 FLOP using benchmark imputation. https://colab.research.google.com/drive/1r3pUMhB7Kh0Gls9eG-v_XefWrye9fVQR?usp=sharing,8.443226,1.866859,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,0.783032,METR-Horizon-v1.1,rechSoSit7zTyOpq0,2023-11-06,4.044959
claude-3-opus-20240229,3.952262,0.294753,2024-02-29,Anthropic,United States of America,,Training compute estimated to be 1.64e25 FLOP from benchmark scores. https://colab.research.google.com/drive/1r3pUMhB7Kh0Gls9eG-v_XefWrye9fVQR?usp=sharing,8.76484,1.706313,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,0.638973,METR-Horizon-v1.1,recyvUC4eTQ5SK6Gr,2024-02-29,4.044959
gpt-4-turbo-2024-04-09,3.732787,0.271786,2024-04-09,OpenAI,United States of America,,Training compute estimated to be 2.2e25 FLOP using benchmark imputation. https://colab.research.google.com/drive/1r3pUMhB7Kh0Gls9eG-v_XefWrye9fVQR?usp=sharing,6.742026,1.938233,,,,0.927933,METR-Horizon-v1.1,rectjHTBu5ytWIXJE,2024-04-09,4.044959
gpt-4o-2024-05-13,6.991195,0.338424,2024-05-13,OpenAI,United States of America,,Training compute estimated to be 3.8e25 FLOP from benchmark scores. https://colab.research.google.com/drive/1r3pUMhB7Kh0Gls9eG-v_XefWrye9fVQR?usp=sharing,12.414331,3.851599,,,,1.267009,METR-Horizon-v1.1,recj2X3OPdvSriHwy,2024-05-13,6.991195
claude-3-5-sonnet-20240620,11.395377,0.401502,2024-06-20,Anthropic,United States of America,2.700000000000001e+25,"Blog post by Dario Amodei includes some info on 3.5 Sonnet compute: https://darioamodei.com/on-deepseek-and-export-controls
""Claude 3.5 Sonnet is a mid-sized model that cost a few $10M's to train (I won't give an exact number). Also, 3.5 Sonnet was not trained in any way that involved a larger or more expensive model (contrary to some rumors).""

Using assumptions about GPU pricing, this lets us estimate compute. https://docs.google.com/spreadsheets/d/1-p-ab6t6dkUM6T7GwnFp85ePTMpZMW7LFY7fW2t8POs/",22.384214,5.489734,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,1.671757,METR-Horizon-v1.1,rec3I5CcPGjF3Riat,2024-06-20,11.395377
o1-preview-2024-09-12,20.326586,0.451441,2024-09-12,OpenAI,United States of America,,,33.379877,11.716193,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,4.420545,METR-Horizon-v1.1,recJhENih6N50V43Q,2024-09-12,20.326586
claude-3-5-sonnet-20241022,20.522872,0.452356,2024-10-22,Anthropic,United States of America,,,40.82028,10.144026,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,2.595677,METR-Horizon-v1.1,recM2JJhzA3vJAZoZ,2024-10-22,20.522872
o1-2024-12-17_medium,38.831588,0.510512,2024-12-17,OpenAI,United States of America,,,64.95249,21.164512,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,7.090121,METR-Horizon-v1.1,recrH7UpKH2XmiFW8,2024-12-17,38.831588
claude-3-7-sonnet-20250219,60.388937,0.558217,2025-02-24,Anthropic,United States of America,3.3499999999999998e+25,https://docs.google.com/spreadsheets/d/10bhwdVrfHI8tysVIz62ZxtvQ30L-HojYvmU18_b-WIM/edit?gid=0#gid=0,107.302719,33.385879,,,,12.09179,METR-Horizon-v1.1,recT0BccE0hbyZxlz,2025-02-24,60.388937
o3-2025-04-16_unknown,119.732634,0.636235,2025-04-16,OpenAI,United States of America,,,191.583353,72.982684,,,,29.981603,METR-Horizon-v1.1,recJSUHrT5GjbcWs8,2025-04-16,119.732634
claude-opus-4-20250514_unknown,100.366123,0.615452,2025-05-22,Anthropic,United States of America,,Flagship model from a leading developer in mid-2025; very likely it used >1e25 FLOP.,163.28087,60.36877,,,,20.429752,METR-Horizon-v1.1,recQ05iSph25jn01x,2025-05-22,119.732634
claude-opus-4-1-20250805,100.472004,0.615713,2025-08-05,Anthropic,United States of America,,,158.015405,60.105845,,,,23.455761,METR-Horizon-v1.1,recwfbaxTfx9WUyj1,2025-08-05,119.732634
gpt-5-2025-08-07_high,203.012577,0.693697,2025-08-07,OpenAI,United States of America,6.6e+25,"Fairly speculative, possibly around 6e25 [CI: 2e25 to 2e26] FLOP. See document below for details

https://docs.google.com/document/d/1V2jIk365LnhH4WDoCw5dYJjZr1Htw8IHaK1noMf5Y48/edit?tab=t.z871imftkus",406.743053,114.211156,,,,38.312431,METR-Horizon-v1.1,recf6lql8VssXVJPx,2025-08-07,203.012577
gemini-3-pro-preview,224.325884,0.709822,2025-11-18,Google DeepMind,United States of America,,Almost certainly >1e25,387.478199,136.865815,,,,54.142849,METR-Horizon-v1.1,rec9ydnu1B0h4VV1V,2025-11-18,224.325884
gpt-5.1-codex-max,223.714694,0.708301,2025-11-19,OpenAI,United States of America,,,396.206723,134.31077,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,50.632499,METR-Horizon-v1.1,recio8mNX4NbLwCsZ,2025-11-19,224.325884
claude-opus-4-5-20251101,292.994594,0.73007,2025-11-24,Anthropic,United States of America,,Flagship model from a leading developer in mid-2025; very likely it used >1e25 FLOP.,638.619623,160.539157,,,,49.430584,METR-Horizon-v1.1,recTbrNrOzFe3rE8x,2025-11-24,292.994594
gpt-5.2-2025-12-11_high,352.249302,0.752875,2025-12-11,OpenAI,United States of America,,,862.339204,191.31908,,,,66.002649,METR-Horizon-v1.1,rect0q40LX3bSJ36F,2025-12-11,352.249302
gpt-5.3-codex,349.530732,0.745439,2026-02-05,OpenAI,United States of America,,,858.325814,192.048863,,,,54.739407,METR-Horizon-v1.1,recLuqsMcBzVzYHTQ,2026-02-05,352.249302
claude-opus-4-6_unknown,718.80683,0.788644,2026-02-05,Anthropic,United States of America,,,3949.750392,319.32091,,,,69.874587,METR-Horizon-v1.1,recF8cs4o1pSE7d15,2026-02-05,718.80683
gemini-3.1-pro-preview,384.147435,0.77035,2026-02-19,Google DeepMind,United States of America,,,694.750898,233.50073,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,,89.801503,METR-Horizon-v1.1,recVAgCABLb9qKDM9,2026-02-19,718.80683
gpt-5.4-2026-03-05_xhigh,341.735276,0.743443,2026-03-05,OpenAI,United States of America,,,768.779526,186.581591,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,,53.877851,METR-Horizon-v1.1,recVtYP9KdMolGtGf,2026-03-05,718.80683
claude-mythos-preview-early,1044.780145,0.85205,2026-04-07,Anthropic,United States of America,,Not disclosed.,3304.261235,508.876789,https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/,,,185.911829,METR-Horizon-v1.1,recW9gDVlpPKxhewo,2026-04-07,1044.780145