METR time horizon
Length of tasks, measured in human working time, that a model completes with 50 percent or 80 percent success. Run on METR's long-task suite.
Time horizon doubling time, as METR publishes it: 129 days (104 to 158) since 2023, 188 days over all time.
No filters. Showing 23 of 23 rows.
1[1–23] AnthropicDossier | No effort set | metr_agents/react | Claude Mythos Preview (early), 50% time horizon: 1,044.8 min (scale 0.0 to 1044.8), 95% confidence interval 508.9 to 3304.31,044.8 min(508.9–3304.3) | 185.9 min97.3 min–398.5 min | 0.9 | Runs unstated | Cost kind unknown | Date not stated |
2[1–23] AnthropicDossierTied #1 | No effort set | metr_agents/react | Opus 4.6, 50% time horizon: 718.8 min (scale 0.0 to 1044.8), 95% confidence interval 316.7 to 3633.8718.8 min(316.7–3633.8) | 69.9 min27 min–170.4 min | 0.8 | Runs unstated | Cost kind unknown | Date not stated |
3[2–9] GoogleDossierTied #1 | No effort set | triframe_inspect/triframe_agent | Gemini 3.1 Pro Preview, 50% time horizon: 384.1 min (scale 0.0 to 1044.8), 95% confidence interval 233.5 to 694.8384.1 min(233.5–694.8) | 89.8 min52 min–158.6 min | 0.8 | Runs unstated | Cost kind unknown | Date not stated |
4[2–14] OpenAIDossierTied #1 | No effort set | triframe_inspect/triframe_agent | GPT-5.2, 50% time horizon: 352.2 min (scale 0.0 to 1044.8), 95% confidence interval 198.1 to 815.2352.2 min(198.1–815.2) | 66 min31.7 min–131.7 min | 0.8 | Runs unstated | Cost kind unknown | Date not stated |
5[2–18] OpenAIDossierTied #1 | No effort set | triframe_inspect/triframe_agent | GPT-5.3 Codex, 50% time horizon: 349.5 min (scale 0.0 to 1044.8), 95% confidence interval 194.9 to 816.4349.5 min(194.9–816.4) | 54.7 min22.4 min–122.1 min | 0.7 | Runs unstated | Cost kind unknown | Date not stated |
6[2–13] OpenAIDossierTied #1 | No effort set | metr_agents/react | GPT-5.4, 50% time horizon: 341.7 min (scale 0.0 to 1044.8), 95% confidence interval 186.6 to 768.8341.7 min(186.6–768.8) | 53.9 min24 min–108.7 min | 0.7 | Runs unstated | Cost kind unknown | Date not stated |
7[2–12] AnthropicDossierTied #1 | No effort set | metr_agents/react | Opus 4.5, 50% time horizon: 293 min (scale 0.0 to 1044.8), 95% confidence interval 161.7 to 623.7293 min(161.7–623.7) | 49.4 min20.7 min–104.6 min | 0.7 | Runs unstated | Cost kind unknown | Date not stated |
8[4–11] GoogleDossierTied #1 | No effort set | metr_agents/react | Gemini 3 Pro Preview, 50% time horizon: 224.3 min (scale 0.0 to 1044.8), 95% confidence interval 139.6 to 379.2224.3 min(139.6–379.2) | 54.1 min25.7 min–103.3 min | 0.7 | Runs unstated | Cost kind unknown | Date not stated |
9[4–13] OpenAIDossierTied #1 | No effort set | triframe_inspect/triframe_agent | GPT-5.1 Codex Max, 50% time horizon: 223.7 min (scale 0.0 to 1044.8), 95% confidence interval 134.3 to 396.2223.7 min(134.3–396.2) | 50.6 min27.3 min–88.4 min | 0.7 | Runs unstated | Cost kind unknown | Date not stated |
10[4–13] OpenAIDossierTied #1 | No effort set | triframe_inspect/triframe_agent | GPT-5, 50% time horizon: 203 min (scale 0.0 to 1044.8), 95% confidence interval 112.6 to 405.6203 min(112.6–405.6) | 38.3 min19.2 min–70.1 min | 0.7 | Runs unstated | Cost kind unknown | Date not stated |
11[8–13] OpenAIDossierTied #1 | No effort set | triframe_inspect/triframe_agent | o3, 50% time horizon: 119.7 min (scale 0.0 to 1044.8), 95% confidence interval 74.6 to 190.9119.7 min(74.6–190.9) | 30 min15.1 min–57.8 min | 0.6 | Runs unstated | Cost kind unknown | Date not stated |
12[8–14] AnthropicDossierTied #1 | No effort set | metr_agents/react | Opus 4.1, 50% time horizon: 100.5 min (scale 0.0 to 1044.8), 95% confidence interval 59.3 to 159.5100.5 min(59.3–159.5) | 23.5 min10.5 min–47.8 min | 0.6 | Runs unstated | Cost kind unknown | Date not stated |
13[8–14] AnthropicDossierTied #1 | No effort set | metr_agents/react | Opus 4, 50% time horizon: 100.4 min (scale 0.0 to 1044.8), 95% confidence interval 60.0 to 163.5100.4 min(60.0–163.5) | 20.4 min8.5 min–41.9 min | 0.6 | Runs unstated | Cost kind unknown | Date not stated |
14[11–15] AnthropicDossierTied #1 | No effort set | metr_agents/react | Claude 3.7 Sonnet, 50% time horizon: 60.4 min (scale 0.0 to 1044.8), 95% confidence interval 33.0 to 104.260.4 min(33.0–104.2) | 12.1 min4.6 min–28.9 min | 0.6 | Runs unstated | Cost kind unknown | Date not stated |
15[13–16] OpenAIDossierTied #1 | No effort set | triframe_inspect/triframe_agent | o1, 50% time horizon: 38.8 min (scale 0.0 to 1044.8), 95% confidence interval 21.2 to 65.038.8 min(21.2–65.0) | 7.1 min3 min–16.6 min | 0.5 | Runs unstated | Cost kind unknown | Date not stated |
16[14–21] AnthropicDossierTied #1 | No effort set | metr_agents/react | Claude 3.5 Sonnet (Oct '24), 50% time horizon: 20.5 min (scale 0.0 to 1044.8), 95% confidence interval 10.1 to 40.820.5 min(10.1–40.8) | 2.6 min0.9 min–7.4 min | 0.5 | Runs unstated | Cost kind unknown | Date not stated |
17[14–18] OpenAIDossierTied #1 | No effort set | duet | o1-preview, 50% time horizon: 20.3 min (scale 0.0 to 1044.8), 95% confidence interval 11.7 to 33.420.3 min(11.7–33.4) | 4.4 min2 min–8.9 min | 0.5 | Runs unstated | Cost kind unknown | Date not stated |
18[16–21] AnthropicDossierTied #1 | No effort set | metr_agents/react | Claude 3.5 Sonnet (June '24), 50% time horizon: 11.4 min (scale 0.0 to 1044.8), 95% confidence interval 5.5 to 22.411.4 min(5.5–22.4) | 1.7 min0.6 min–4.6 min | 0.4 | Runs unstated | Cost kind unknown | Date not stated |
19[17–22] OpenAIDossierTied #1 | No effort set | metr_agents/react | GPT-4o (May '24), 50% time horizon: 7 min (scale 0.0 to 1044.8), 95% confidence interval 4.0 to 12.97 min(4.0–12.9) | 1.3 min0.6 min–3 min | 0.3 | Runs unstated | Cost kind unknown | Date not stated |
20[18–23] OpenAIDossierTied #1 | No effort set | metr_agents/react | GPT-4 Turbo Preview (1106), 50% time horizon: 4 min (scale 0.0 to 1044.8), 95% confidence interval 1.9 to 8.44 min(1.9–8.4) | 0.8 min0.3 min–2.4 min | 0.3 | Runs unstated | Cost kind unknown | Date not stated |
21[18–23] OpenAIDossierTied #1 | No effort set | modular-public | GPT-4 (0314), 50% time horizon: 4 min (scale 0.0 to 1044.8), 95% confidence interval 1.9 to 8.04 min(1.9–8.0) | 0.9 min0.3 min–2.5 min | 0.3 | Runs unstated | Cost kind unknown | Date not stated |
22[18–23] AnthropicDossierTied #1 | No effort set | metr_agents/react | Claude 3 Opus, 50% time horizon: 4 min (scale 0.0 to 1044.8), 95% confidence interval 1.7 to 8.84 min(1.7–8.8) | 0.6 min0.2 min–2.1 min | 0.3 | Runs unstated | Cost kind unknown | Date not stated |
23[18–23] OpenAIDossierTied #1 | No effort set | metr_agents/react | GPT-4 Turbo, 50% time horizon: 3.7 min (scale 0.0 to 1044.8), 95% confidence interval 2.0 to 6.73.7 min(2.0–6.7) | 0.9 min0.4 min–2.2 min | 0.3 | Runs unstated | Cost kind unknown | Date not stated |
- 1AnthropicDossier
metr_agents/react
50% time horizon
Claude Mythos Preview (early), 50% time horizon: 1,044.8 min (scale 0.0 to 1044.8), 95% confidence interval 508.9 to 3304.31,044.8 min(508.9–3304.3)- 80% time horizon
- 185.9 min
- Average score
- 0.9
50% time horizon
Opus 4.6, 50% time horizon: 718.8 min (scale 0.0 to 1044.8), 95% confidence interval 316.7 to 3633.8718.8 min(316.7–3633.8)- 80% time horizon
- 69.9 min
- Average score
- 0.8
50% time horizon
Gemini 3.1 Pro Preview, 50% time horizon: 384.1 min (scale 0.0 to 1044.8), 95% confidence interval 233.5 to 694.8384.1 min(233.5–694.8)- 80% time horizon
- 89.8 min
- Average score
- 0.8
50% time horizon
GPT-5.2, 50% time horizon: 352.2 min (scale 0.0 to 1044.8), 95% confidence interval 198.1 to 815.2352.2 min(198.1–815.2)- 80% time horizon
- 66 min
- Average score
- 0.8
50% time horizon
GPT-5.3 Codex, 50% time horizon: 349.5 min (scale 0.0 to 1044.8), 95% confidence interval 194.9 to 816.4349.5 min(194.9–816.4)- 80% time horizon
- 54.7 min
- Average score
- 0.7
50% time horizon
GPT-5.4, 50% time horizon: 341.7 min (scale 0.0 to 1044.8), 95% confidence interval 186.6 to 768.8341.7 min(186.6–768.8)- 80% time horizon
- 53.9 min
- Average score
- 0.7
50% time horizon
Opus 4.5, 50% time horizon: 293 min (scale 0.0 to 1044.8), 95% confidence interval 161.7 to 623.7293 min(161.7–623.7)- 80% time horizon
- 49.4 min
- Average score
- 0.7
50% time horizon
Gemini 3 Pro Preview, 50% time horizon: 224.3 min (scale 0.0 to 1044.8), 95% confidence interval 139.6 to 379.2224.3 min(139.6–379.2)- 80% time horizon
- 54.1 min
- Average score
- 0.7
50% time horizon
GPT-5.1 Codex Max, 50% time horizon: 223.7 min (scale 0.0 to 1044.8), 95% confidence interval 134.3 to 396.2223.7 min(134.3–396.2)- 80% time horizon
- 50.6 min
- Average score
- 0.7
50% time horizon
GPT-5, 50% time horizon: 203 min (scale 0.0 to 1044.8), 95% confidence interval 112.6 to 405.6203 min(112.6–405.6)- 80% time horizon
- 38.3 min
- Average score
- 0.7
50% time horizon
o3, 50% time horizon: 119.7 min (scale 0.0 to 1044.8), 95% confidence interval 74.6 to 190.9119.7 min(74.6–190.9)- 80% time horizon
- 30 min
- Average score
- 0.6
50% time horizon
Opus 4.1, 50% time horizon: 100.5 min (scale 0.0 to 1044.8), 95% confidence interval 59.3 to 159.5100.5 min(59.3–159.5)- 80% time horizon
- 23.5 min
- Average score
- 0.6
50% time horizon
Opus 4, 50% time horizon: 100.4 min (scale 0.0 to 1044.8), 95% confidence interval 60.0 to 163.5100.4 min(60.0–163.5)- 80% time horizon
- 20.4 min
- Average score
- 0.6
50% time horizon
Claude 3.7 Sonnet, 50% time horizon: 60.4 min (scale 0.0 to 1044.8), 95% confidence interval 33.0 to 104.260.4 min(33.0–104.2)- 80% time horizon
- 12.1 min
- Average score
- 0.6
50% time horizon
o1, 50% time horizon: 38.8 min (scale 0.0 to 1044.8), 95% confidence interval 21.2 to 65.038.8 min(21.2–65.0)- 80% time horizon
- 7.1 min
- Average score
- 0.5
50% time horizon
Claude 3.5 Sonnet (Oct '24), 50% time horizon: 20.5 min (scale 0.0 to 1044.8), 95% confidence interval 10.1 to 40.820.5 min(10.1–40.8)- 80% time horizon
- 2.6 min
- Average score
- 0.5
50% time horizon
o1-preview, 50% time horizon: 20.3 min (scale 0.0 to 1044.8), 95% confidence interval 11.7 to 33.420.3 min(11.7–33.4)- 80% time horizon
- 4.4 min
- Average score
- 0.5
50% time horizon
Claude 3.5 Sonnet (June '24), 50% time horizon: 11.4 min (scale 0.0 to 1044.8), 95% confidence interval 5.5 to 22.411.4 min(5.5–22.4)- 80% time horizon
- 1.7 min
- Average score
- 0.4
50% time horizon
GPT-4o (May '24), 50% time horizon: 7 min (scale 0.0 to 1044.8), 95% confidence interval 4.0 to 12.97 min(4.0–12.9)- 80% time horizon
- 1.3 min
- Average score
- 0.3
50% time horizon
GPT-4 Turbo Preview (1106), 50% time horizon: 4 min (scale 0.0 to 1044.8), 95% confidence interval 1.9 to 8.44 min(1.9–8.4)- 80% time horizon
- 0.8 min
- Average score
- 0.3
50% time horizon
GPT-4 (0314), 50% time horizon: 4 min (scale 0.0 to 1044.8), 95% confidence interval 1.9 to 8.04 min(1.9–8.0)- 80% time horizon
- 0.9 min
- Average score
- 0.3
50% time horizon
Claude 3 Opus, 50% time horizon: 4 min (scale 0.0 to 1044.8), 95% confidence interval 1.7 to 8.84 min(1.7–8.8)- 80% time horizon
- 0.6 min
- Average score
- 0.3
50% time horizon
GPT-4 Turbo, 50% time horizon: 3.7 min (scale 0.0 to 1044.8), 95% confidence interval 2.0 to 6.73.7 min(2.0–6.7)- 80% time horizon
- 0.9 min
- Average score
- 0.3