BrokenArXiv leaderboard
As of 16 Nov 2026, the top score on BrokenArXiv is 84.4% by Gemini 4 Carbon (Google), ahead of GPT-6.1 Sol at 83.1%, across 8 model configurations with a published result (MathArena). Statements from recent arXiv papers altered slightly so they are false: a model scores when it points out the flaw instead of 'proving' it. MathArena. Shown for reference; few models so far. BrokenArXiv is shown for reference only and does not count toward the BenchLeader Index.
Key facts (as of 16 Nov 2026)
- Gemini 4 Carbon holds the top BrokenArXiv score at 84.4% (MathArena, 16 Nov 2026).
- GPT-6.1 Sol is second on BrokenArXiv at 83.1% (MathArena, 16 Nov 2026).
- BrokenArXiv: Statements from recent arXiv papers altered slightly so they are false: a model scores when it points out the flaw instead of 'proving' it. MathArena. Shown for reference; few models so far (MathArena).
- 8 model configurations have a published BrokenArXiv result on BenchLeader (BenchLeader, 16 Nov 2026).
- BrokenArXiv does not feed the BenchLeader Index; it is shown for reference (BenchLeader methodology).
Statements from recent arXiv papers altered slightly so they are false: a model scores when it points out the flaw instead of 'proving' it. MathArena. Shown for reference; few models so far.
- Published by
- MathArena
- Category
- Maths
- Index weight
- Reference only
- Models
- 8
- Data as of
- 16 Nov 2026
MIT — eth-sri/matharena.
- 1Gemini 4 Carbon (high)84.4%
- 2GPT-6.1 Sol (max)83.1%
- 3Claude Opus 5.5 (high)77.3%
- 4Inkling 2 Small63.2%
- 5Kimi K3.1 Pro (max)61.9%
- 6DeepSeek V4.1 Flash 1017 (max)48.7%
- 7Gemini 4 Argon (high)47.9%
- 8Muse Spark Open (xhigh)30.1%
| # | |||||
|---|---|---|---|---|---|
| 1 | 84.4% | 68.9 | $16.35 | 540k | |
| 2 | 83.1% | 66.6 | $1.14 | 80k | |
| 3 | 77.3% | 66.0 | $2.66 | 86k | |
| 4 | 63.2% | 57.9 | $0.220 | 128k | |
| 5 | 61.9% | 59.9 | $2.35 | 125k | |
| 6 | 48.7% | 57.7 | $0.290 | 184k | |
| 7 | 47.9% | 65.3 | $9.55 | 524k | |
| 8 | 30.1% | 53.5 | $0.450 | 56k |
AI reasoning and math benchmark leaderboard · All LLM benchmarks
Related benchmarks
Top head-to-heads on BrokenArXiv
Cite as: BenchLeader, “BrokenArXiv leaderboard”, data as of 16 Nov 2026, https://www.benchleader.com/benchmarks/matharena_brokenarxiv.
BrokenArXiv: questions
- What does BrokenArXiv measure?
- Statements from recent arXiv papers altered slightly so they are false: a model scores when it points out the flaw instead of 'proving' it. MathArena. Shown for reference; few models so far. Scores are reported in percent of tasks solved; higher is better.
- What is the top score on BrokenArXiv?
- As of 16 Nov 2026, the highest score on BrokenArXiv is 84.4% by Gemini 4 Carbon, ahead of GPT-6.1 Sol at 83.1%. Scores come from MathArena and are re-read daily.
- Which AI model leads BrokenArXiv?
- Gemini 4 Carbon leads BrokenArXiv with 84.4% as of 16 Nov 2026, ahead of GPT-6.1 Sol at 83.1%.
- How many models have BrokenArXiv results?
- 8 model configurations have a BrokenArXiv result on BenchLeader, all taken from MathArena.
- Who runs BrokenArXiv and how often is it updated?
- BrokenArXiv is published by MathArena. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does BrokenArXiv count toward the BenchLeader Index?
- No. BrokenArXiv is shown for reference but left out of the composite index.