Benchmark profile

Massive Multi-discipline Multimodal Understanding Pro (MMMU-Pro)

A harder multimodal benchmark for frontier models that combines text with images, diagrams, charts, and academic visual reasoning tasks.

Data verified July 23, 2026

Top models on MMMU-Pro — July 23, 2026

As of July 23, 2026, GPT-5.4 Pro leads the MMMU-Pro leaderboard with 94% , followed by Gemini 3.1 Pro (83.9%) and Gemini 3.5 Flash (83.6%).

1Closed

GPT-5.4 Pro

OpenAI

gpt-5-4-pro

94%

Overall 60.89Context 1.05M

2Closed

Gemini 3.1 Pro

Google

gemini-3-1-pro

83.9%

Overall 55.3Context 1M

3Closed

Gemini 3.5 Flash

Google

gemini-3-5-flash

83.6%

Overall 64.75Context 1M

34 modelsMultimodal & Grounded45% of category scoreRefreshingUpdated July 23, 2026

Leaderboard (34 models)

Score

GPT-5.4 ProOpenAI · Closed

94%

Gemini 3.1 ProGoogle · Closed

83.9%

Gemini 3.5 FlashGoogle · Closed

83.6%

GPT-5.6 SolOpenAI · Closed

83%

Kimi K3Moonshot AI · Closed

81.6%

GPT-5.5OpenAI · Closed

81.2%

GPT-5.4OpenAI · Closed

81.2%

Gemini 3 ProGoogle · Closed

81%

GPT-5.6 TerraOpenAI · Closed

80.7%

Muse SparkMeta · Closed

80.4%

GPT-5.2OpenAI · Closed

79.5%

Kimi K2.6Moonshot AI · Open weight

79.4%

Qwen3.7 PlusAlibaba · Closed

79%

Qwen3.5 397BAlibaba · Open weight

79%

Qwen3.6 PlusAlibaba · Closed

78.8%

Kimi K2.5Moonshot AI · Open weight

78.5%

Kimi K2.5 (Reasoning)Moonshot AI · Closed

78.5%

GPT-5.6 LunaOpenAI · Closed

78.4%

MiniMax M3MiniMax · Open weight

78.1%

Grok 4.3xAI · Closed

78.1%

MiMo-V2.5Xiaomi · Closed

77.9%

Claude Opus 4.6Anthropic · Closed

77.3%

Gemma 4 31BGoogle · Open weight

76.9%

GPT-5.4 miniOpenAI · Closed

76.6%

Qwen3.6-27BAlibaba · Open weight

75.8%

Qwen3.6-35B-A3BAlibaba · Open weight

75.3%

Grok 4.20xAI · Closed

75.2%

Gemma 4 26B A4BGoogle · Open weight

73.8%

InklingThinking Machines Lab · Open weight

73.5%

Interfaze BetaInterfaze · Closed

71.1%

Claude Opus 4.5Anthropic · Closed

70.6%

Gemma 4 12BGoogle · Open weight

69.1%

GPT-5.4 nanoOpenAI · Closed

66.1%

Command A+Cohere · Open weight

63%

According to BenchLM.ai, GPT-5.4 Pro leads the MMMU-Pro benchmark with a score of 94%, followed by Gemini 3.1 Pro (83.9%) and Gemini 3.5 Flash (83.6%). The scores show moderate spread, with meaningful differences between the top tier and mid-tier models.

34 models have been evaluated on MMMU-Pro. The benchmark falls in the Multimodal & Grounded category. This category carries a 12% weight in BenchLM.ai's overall scoring system. Within that category, MMMU-Pro contributes 45% of the category score, so strong performance here directly affects a model's overall ranking.

About MMMU-Pro

Year

2024

Tasks

Multimodal academic reasoning

Format

Image + text question answering

Difficulty

Frontier multimodal

MMMU-Pro extends the original MMMU setup with more difficult multimodal questions and stronger separation at the top end of the model market.

MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark

BenchLM freshness & provenance

Version

MMMU-Pro 2024

Refresh cadence

Annual

Staleness state

Refreshing

Question availability

Public benchmark set

Refreshing

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

FAQ

What does MMMU-Pro measure?

A harder multimodal benchmark for frontier models that combines text with images, diagrams, charts, and academic visual reasoning tasks.

Which model scores highest on MMMU-Pro?

GPT-5.4 Pro by OpenAI currently leads with a score of 94% on MMMU-Pro.

How many models are evaluated on MMMU-Pro?

34 AI models have been evaluated on MMMU-Pro on BenchLM.

Compare Top Models on MMMU-Pro

GPT-5.4 Pro vs Gemini 3.1 Pro Gemini 3.1 Pro vs Gemini 3.5 Flash Gemini 3.5 Flash vs GPT-5.6 Sol GPT-5.6 Sol vs Kimi K3

Last updated: July 23, 2026 · BenchLM version MMMU-Pro 2024

Choose a model with this week’s evidence

Join 2,000+ readers for ranking moves, pricing changes, and the claims that still need proof.

One email each week. Unsubscribe anytime.