METR 50% Time Horizon
Length of software and research tasks, measured in how long they take skilled humans, that the model completes with 50% reliability (METR-Horizon v1.1). Shares weight with the Epoch copy of the METR task suite.
| Category | Agents |
|---|---|
| Metric | Minutes at 50% success |
| Source | METR (original) |
| License | None stated |
| Benchmark page | metr.org |
| Status | Scored |
| Weight | 1.2 within Agents |
| Fit | Discrimination 0.87, noise 0.24 |
| Lineage | Shares weight with metr_time_horizons |
Results
| # | Model | Provider | Result | Date |
|---|---|---|---|---|
| 1 | Claude Mythos Preview | Anthropic | $1,045 | 2026-11-11 |
| 2 | Claude Opus 4.6 | Anthropic | $719 | 2026-11-11 |
| 3 | Gemini 3.1 Pro Preview | $384 | 2026-11-11 | |
| 4 | GPT-5.2 | OpenAI | $352 | 2026-11-11 |
| 5 | GPT-5.3 Codex | OpenAI | $350 | 2026-11-11 |
| 6 | GPT-5.4 | OpenAI | $342 | 2026-11-11 |
| 7 | Claude Opus 4.5 | Anthropic | $293 | 2026-11-11 |
| 8 | Gemini 3 Pro Preview | $224 | 2026-11-11 | |
| 9 | GPT-5.1 Codex Max | OpenAI | $224 | 2026-11-11 |
| 10 | GPT-5 | OpenAI | $203 | 2026-11-11 |
| 11 | o3 | OpenAI | $120 | 2026-11-11 |
| 12 | Claude Opus 4.1 | Anthropic | $100 | 2026-11-11 |
| 13 | Claude Opus 4 | Anthropic | $100 | 2026-11-11 |
| 14 | Claude Sonnet 3.7 | Anthropic | $60 | 2026-11-11 |
| 15 | o1 | OpenAI | $39 | 2026-11-11 |
| 16 | Claude Sonnet 3.5 v2 | Anthropic | $21 | 2026-11-11 |
| 17 | o1-preview | OpenAI | $20 | 2026-11-11 |
| 18 | Claude 3.5 Sonnet | Anthropic | $11 | 2026-11-11 |
| 19 | GPT-4o | OpenAI | $7 | 2026-11-11 |
| 20 | GPT-4 | OpenAI | $4 | 2026-11-11 |
| 21 | Claude 3 Opus | Anthropic | $4 | 2026-11-11 |
| 22 | GPT-4 Turbo | OpenAI | $4 | 2026-11-11 |
| 23 | GPT-3.5-turbo | OpenAI | $1 | 2026-11-11 |
| 24 | davinci-002 | OpenAI | $0 | 2026-11-11 |