ForecastBench
Forecasting real future events, scored as they resolve.
As of 19 Sept 2026, o3 leads ForecastBench on BenchLeader with 62.5%, ahead of Claude Opus 4.1 at 62.0%, across 73 model configurations with a published result.
- Published by
- ForecastBenchdata via Epoch AI Benchmarking Hub
- Category
- Reasoning
- Index weight
- Reference only
- Models
- 73
- Data as of
- 19 Sept 2026
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
Models forecast real-world questions; scores accrue as the questions resolve, compared with superforecasters.
How it is scored
Overall Brier-based score, as published by ForecastBench.
What to keep in mind
Scores change as events resolve; a lower Brier score is better and the site rescales it.
- 1o362.5%
- 2Claude Opus 4.162.0%
- 3Claude Sonnet 4.561.9%
- 4o4-mini61.8%
- 5Claude 3.7 Sonnet61.8%
- 6GPT-4.561.7%
- 7GPT-4.161.5%
- 8GPT-561.4%
- 9Claude Haiku 4.561.4%
- 10Grok 4.2061.4%
- 11Gemini 2.5 Pro61.3%
- 12Gemini 3 Pro61.2%
- 13Claude Sonnet 4.661.2%
- 14GPT-5.561.1%
- 15Claude Opus 461.1%
73 of 73
| # | ||||
|---|---|---|---|---|
| 1 |  o3OpenAI | 62.5% | 60.3 | 2025-04-16 |
| 2 |  Claude Opus 4.1Anthropic | 62.0% | 51.9 | 2025-08-05 |
| 3 | Claude Sonnet 4.5Anthropic | 61.9% | 54.3 | 2025-09-29 |
| 4 | o4-miniOpenAI | 61.8% | 57.4 | 2025-04-16 |
| 5 | Claude 3.7 SonnetAnthropic | 61.8% | 50.2 | 2025-02-24 |
| 6 | GPT-4.5OpenAI | 61.7% | 50.2 | 2025-02-27 |
| 7 | GPT-4.1OpenAI | 61.5% | 49.0 | 2025-04-14 |
| 8 | GPT-5OpenAI | 61.4% | 57.2 | 2025-08-07 |
| 9 | Claude Haiku 4.5Anthropic | 61.4% | 48.1 | 2025-10-15 |
| 10 |  Grok 4.20xAI | 61.4% | – | 2026-02-17 |
| 11 |  Gemini 2.5 ProGoogle | 61.3% | 54.5 | 2025-03-31 |
| 12 | Gemini 3 ProGoogle | 61.2% | 61.1 | 2025-11-18 |
| 13 | Claude Sonnet 4.6Anthropic | 61.2% | 59.0 | 2026-02-17 |
| 14 | GPT-5.5OpenAI | 61.1% | 63.2 | 2026-04-23 |
| 15 | Claude Opus 4Anthropic | 61.1% | 52.9 | 2025-05-22 |
| 16 | GPT-5 miniOpenAI | 61.0% | 56.8 | 2025-08-07 |
| 17 | Grok 4.1thinkingxAI | 61.0% | 56.0 | 2025-11-19 |
| 18 |  GLM-5Zhipu AIopen | 61.0% | 54.7 | 2026-02-11 |
| 19 | Grok 4xAI | 60.9% | 58.5 | 2025-07-09 |
| 20 | Grok 4.20thinkingxAI | 60.7% | 59.9 | 2026-02-17 |
| 21 | Claude Opus 4.5Anthropic | 60.7% | 58.2 | 2025-11-24 |
| 22 | Claude 3.5 SonnetAnthropic | 60.7% | 46.6 | 2024-10-22 |
| 23 | Gemini 2.5 FlashGoogle | 60.6% | 52.5 | 2025-04-17 |
| 24 | Grok 4 FastxAI | 60.5% | 54.3 | 2025-09-19 |
| 25 |  MiniMax-M3MiniMaxopen ↗ | 60.4% | 57.5 | 2026-06-01 |
| 26 | Grok 4.3xAI | 60.4% | 50.5 | 2026-04-17 |
| 27 |  Kimi K2Moonshot AIopen | 60.2% | 51.0 | 2025-07-12 |
| 28 | Claude Sonnet 4Anthropic | 60.2% | 49.7 | 2025-05-22 |
| 29 | GPT-5.2OpenAI | 60.1% | 58.4 | 2025-12-11 |
| 30 | Claude Opus 4.6Anthropic | 60.0% | 63.7 | 2026-02-05 |
| 31 |  DeepSeek R1DeepSeekopen | 60.0% | 48.7 | 2025-01-20 |
| 32 | Claude Sonnet 5Anthropic | 59.9% | 57.2 | 2026-06-30 |
| 33 |  Llama 3.1 405BMetaopen | 59.9% | 43.5 | 2024-07-23 |
| 34 | Kimi K2 0905Moonshot AIopen | 59.8% | 51.3 | 2025-09-05 |
| 35 |  Qwen3 235B A22BAlibabaopen | 59.7% | 49.6 | 2025-04-29 |
| 36 | Claude Opus 4.7Anthropic | 59.6% | 64.5 | 2026-04-16 |
| 37 | o3-miniOpenAI | 59.6% | 50.4 | 2025-01-31 |
| 38 | GPT-5.4OpenAI | 59.5% | 59.2 | 2026-03-05 |
| 39 | GPT-4 TurboOpenAI | 59.4% | 41.5 | 2024-04-09 |
| 40 | Claude Opus 4.8Anthropic | 59.3% | 61.9 | 2026-05-28 |
| 41 | Gemini 3.5 FlashGoogle | 59.2% | 51.3 | 2026-05-19 |
| 42 | GLM-4.5-AirZhipu AIopen | 59.2% | 47.2 | 2025-07-20 |
| 43 | GPT-5 nanoOpenAI | 59.1% | 55.9 | 2025-08-07 |
| 44 | DeepSeek V3DeepSeekopen | 59.1% | 44.8 | 2024-12-26 |
| 45 | Gemini 3.1 ProGoogle | 59.0% | 63.9 | 2026-02-19 |
| 46 | Llama 3.3 70BMetaopen | 58.6% | 41.4 | 2024-12-06 |
| 47 | Llama 3-8BMetaopen | 58.6% | 35.6 | 2024-04-18 |
| 48 | Gemini 3 FlashGoogle | 58.5% | 57.7 | 2025-12-17 |
| 49 | Gemini 1.5 Pro 001Google | 58.5% | 45.1 | 2024-05-14 |
| 50 | Claude 3 OpusAnthropic | 58.4% | 39.6 | 2024-02-29 |
| 51 | QwQ-32BAlibabaopen | 58.3% | 46.1 | 2024-11-28 |
| 52 | GPT-5.1OpenAI | 58.1% | 57.2 | 2025-11-13 |
| 53 | GPT 4 0613OpenAI | 57.8% | 40.2 | 2023-06-13 |
| 54 | GPT-4oOpenAI | 57.7% | 44.3 | 2024-05-13 |
| 55 | Qwen1 5 110BAlibabaopen | 57.7% | 42.7 | 2024-04-25 |
| 56 | Llama 4 MaverickMetaopen | 57.5% | 43.4 | 2025-04-05 |
| 57 | Qwen2.5 72BAlibabaopen | 57.5% | 42.7 | 2024-09-19 |
| 58 | Llama 4 ScoutMetaopen | 57.5% | 40.2 | 2025-04-05 |
| 59 | GPT-5.4 nanoOpenAI | 57.3% | – | 2026-03-17 |
| 60 |  Mistral Large 2Mistral AIopen | 57.1% | 41.0 | 2024-07-24 |
Cite as: BenchLeader, “ForecastBench leaderboard”, https://www.benchleader.com/benchmarks/forecastbench, data as of 19 Sept 2026.
ForecastBench: questions
- What does ForecastBench measure?
- Models forecast real-world questions; scores accrue as the questions resolve, compared with superforecasters. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads ForecastBench?
- o3 leads ForecastBench with 62.5% as of 19 Sept 2026, ahead of Claude Opus 4.1 at 62.0%.
- How many models have ForecastBench results?
- 73 model configurations have a ForecastBench result on BenchLeader, all taken from ForecastBench via Epoch AI Benchmarking Hub.
- Who runs ForecastBench and how often is it updated?
- ForecastBench is published by ForecastBench. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does ForecastBench count toward the BenchLeader Index?
- No. ForecastBench is shown for reference but left out of the composite index.