WeirdML v3 leaderboard
As of 16 Nov 2026, the top score on WeirdML v3 is 42.2% by GPT-6 Astra (OpenAI), ahead of GPT-6.1 Sol at 38.3%, across 26 model configurations with a published result (WeirdML, result dated 3 Sep 2026). Eleven unusual machine-learning tasks solved by a coding agent that writes and runs its own training code. The leaderboard's official score rewards reaching a good result in few tokens as well as the result itself; not comparable with WeirdML v2. WeirdML v3 is shown for reference only and does not count toward the BenchLeader Index.
Key facts (as of 16 Nov 2026)
- GPT-6 Astra holds the top WeirdML v3 score at 42.2% (WeirdML, 3 Sep 2026).
- GPT-6.1 Sol is second on WeirdML v3 at 38.3% (WeirdML, 29 Sep 2026).
- WeirdML v3: Eleven unusual machine-learning tasks solved by a coding agent that writes and runs its own training code. The leaderboard's official score rewards reaching a good result in few tokens as well as the result itself; not comparable with WeirdML v2 (WeirdML).
- 26 model configurations have a published WeirdML v3 result on BenchLeader (BenchLeader, 16 Nov 2026).
- WeirdML v3 does not feed the BenchLeader Index; it is shown for reference (BenchLeader methodology).
Eleven unusual machine-learning tasks solved by a coding agent that writes and runs its own training code. The leaderboard's official score rewards reaching a good result in few tokens as well as the result itself; not comparable with WeirdML v2.
- Published by
- WeirdMLdata via Epoch AI Benchmarking Hub
- Category
- Coding
- Index weight
- Reference only
- Models
- 26
- Data as of
- 16 Nov 2026
- Newest result
- 26 Oct 2026
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
- 1GPT-6 Astra (xhigh)42.2%
- 2GPT-6.1 Sol (xhigh)38.3%
- 3Claude Opus 5.5 (xhigh)31.2%
- 4GPT-6.1 Terra (xhigh)28.1%
- 5Claude Fable 5.1 (xhigh)26.0%
- 6Gemini 4 Argon (high)20.3%
- 7Claude Sonnet 5.5 (xhigh)19.9%
- 8GPT-6 Sol (xhigh)19.7%
- 9Claude Opus 5 (xhigh)19.0%
- 10Claude Fable 5 (xhigh)15.6%
- 11Muse Wonder (xhigh)15.2%
- 12GPT-5.6 Sol (xhigh)15.0%
- 13Claude Haiku 5.5 (xhigh)11.7%
- 14Kimi K3.1 Pro (xhigh)9.9%
- 15DeepSeek V4.1 Flash 1017 (high)9.2%
| # | ||||
|---|---|---|---|---|
| 1 | 42.2% | 66.0 | 2026-09-03 | |
| 2 | 38.3% | 63.6 | 2026-09-29 | |
| 3 | 31.2% | 65.4 | 2026-09-22 | |
| 4 | 28.1% | 60.0 | 2026-10-26 | |
| 5 | 26.0% | 64.6 | 2026-09-01 | |
| 6 | 20.3% | 65.3 | 2026-10-13 | |
| 7 | 19.9% | 62.5 | 2026-09-28 | |
| 8 | 19.7% | 58.4 | 2026-09-22 | |
| 9 | 19.0% | 60.1 | 2026-07-24 | |
| 10 | 15.6% | – | 2026-06-09 | |
| 11 | 15.2% | – | 2026-10-12 | |
| 12 | 15.0% | 63.9 | 2026-07-09 | |
| 13 | 11.7% | 56.7 | 2026-10-07 | |
| 14 | 9.9% | – | 2026-10-23 | |
| 15 | 9.2% | 51.2 | 2026-10-17 | |
| 16 | 8.4% | 57.4 | 2026-09-21 | |
| 17 | 7.9% | 52.0 | 2026-09-22 | |
| 18 | 7.5% | – | 2026-07-16 | |
| 19 | 7.4% | 59.4 | 2026-09-02 | |
| 20 | 6.7% | 58.1 | 2026-09-02 | |
| 21 | 6.5% | 59.7 | 2025-11-24 | |
| 22 | 6.0% | 55.6 | 2026-09-09 | |
| 23 | 5.2% | 58.2 | 2026-07-09 | |
| 24 | 5.0% | 59.9 | 2025-12-11 | |
| 25 | 4.5% | 48.3 | 2026-10-06 | |
| 26 | 3.9% | 50.0 | 2026-07-15 |
AI coding benchmark leaderboard · All LLM benchmarks
Related benchmarks
Top head-to-heads on WeirdML v3
Cite as: BenchLeader, “WeirdML v3 leaderboard”, data as of 16 Nov 2026, https://www.benchleader.com/benchmarks/weirdml_v3.
WeirdML v3: questions
- What does WeirdML v3 measure?
- Eleven unusual machine-learning tasks solved by a coding agent that writes and runs its own training code. The leaderboard's official score rewards reaching a good result in few tokens as well as the result itself; not comparable with WeirdML v2. Scores are reported in percent of tasks solved; higher is better.
- What is the top score on WeirdML v3?
- As of 16 Nov 2026, the highest score on WeirdML v3 is 42.2% by GPT-6 Astra (result dated 3 Sep 2026), ahead of GPT-6.1 Sol at 38.3%. Scores come from WeirdML and are re-read daily.
- Which AI model leads WeirdML v3?
- GPT-6 Astra leads WeirdML v3 with 42.2% as of 16 Nov 2026 (result dated 3 Sep 2026), ahead of GPT-6.1 Sol at 38.3%.
- How many models have WeirdML v3 results?
- 26 model configurations have a WeirdML v3 result on BenchLeader, all taken from WeirdML via Epoch AI Benchmarking Hub.
- Who runs WeirdML v3 and how often is it updated?
- WeirdML v3 is published by WeirdML. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does WeirdML v3 count toward the BenchLeader Index?
- No. WeirdML v3 is shown for reference but left out of the composite index.