METR Time Horizons
How long a task (in human expert time) the model completes with 50% reliability. Scored on METR’s task-suite average.
Agent Index weight
10%
in the Long-horizon autonomy pillar
Models tested
35
from the benchmark’s leaderboard and reports
Results
| # | Model | Score | Normalised | Harness | Cost/task | Date |
|---|---|---|---|---|---|---|
| 1 | Claude Mythos Preview Anthropic | 85.2% | source | — | — | |
| 2 | Claude Opus 4.6 Supported Anthropic | 78.9% | source | — | — | |
| 3 | Gemini 3.1 Pro Preview Supported Google | 77.0% | source | — | — | |
| 4 | GPT-5.2 Supported OpenAI | 75.3% | source | — | — | |
| 5 | Claude Opus 4.5 (latest) Supported Anthropic | 75.0% | source | — | — | |
| 6 | GPT-5.3 Codex Supported OpenAI | 74.5% | source | — | — | |
| 7 | GPT-5.4 Supported OpenAI | 74.3% | source | — | — | |
| 8 | Gemini 3 Pro Preview Supported Google | 71.0% | source | — | — | |
| 9 | GPT-5.1 Codex Max Estimated OpenAI | 70.8% | source | — | — | |
| 10 | GPT-5 Supported OpenAI | 69.6% | source | — | — | |
| 11 | Claude Sonnet 4.5 (latest) Supported Anthropic | 67.4% | source | — | — | |
| 12 | Claude Opus 4.1 (latest) Supported Anthropic | 66.8% | source | — | — | |
| 13 | Grok 4 Supported xAI | 66.6% | source | — | — | |
| 14 | o3 Supported OpenAI | 65.4% | source | — | — | |
| 15 | Claude-Opus-4 Supported Anthropic | 63.9% | source | — | — | |
| 16 | o4-mini Estimated OpenAI | 63.9% | source | — | — | |
| 17 | Claude Sonnet 4 Supported Anthropic | 62.0% | source | — | — | |
| 18 | Claude 3.7 Sonnet Supported Anthropic | 60.0% | source | — | — | |
| 19 | Kimi K2 Thinking Estimated Moonshot AI | 59.2% | source | — | — | |
| 20 | GPT OSS 120B Supported OpenAI | 56.6% | source | — | — | |
| 21 | Gemini 2.5 Pro Supported Google | 55.4% | source | — | — | |
| 22 | DeepSeek-R1 Estimated DeepSeek | 53.8% | source | — | — | |
| 23 | o1 Estimated OpenAI | 51.1% | source | — | — | |
| 24 | DeepSeek V3 0324 DeepSeek | 49.6% | source | — | — | |
| 25 | DeepSeek-V3 DeepSeek | 47.4% | source | — | — | |
| 26 | Claude-Sonnet-3.5 Estimated Anthropic | 45.2% | source | — | — | |
| 27 | GPT-4o (2024-08-06) Supported OpenAI | 40.8% | source | — | — | |
| 28 | GPT-4 Turbo OpenAI | 36.7% | source | — | — | |
| 29 | GPT-4 OpenAI | 36.1% | source | — | — | |
| 30 | Qwen2.5-72B Estimated Alibaba (Qwen) | 35.8% | source | — | — | |
| 31 | Qwen2-72B Alibaba (Qwen) | 29.9% | source | — | — | |
| 32 | Claude 3 Opus Estimated Anthropic | 29.5% | source | — | — | |
| 33 | GPT-3.5 Turbo Instruct OpenAI | 21.5% | source | — | — | |
| 34 | GPT-3 175B (davinci) OpenAI | 16.2% | source | — | — | |
| 35 | GPT-2 (1.5B) OpenAI | 10.1% | source | — | — |
For models without a METR Time Horizons result, the index estimates one from Epoch’s Capabilities Index using a logistic fit on 29 models (residual SD 0.13 on the logit scale). Estimates are labelled and widen the model’s range.