live
weekly refresh
basedagi.org
benchmark evidence

MMLU-Pro

Massive Multitask Language Understanding — Professional. 14 subjects, 5-shot CoT.

winner on MMLU-Pro
direct benchmark result, not a broad vertical composite | source row dated 2025-05-01
scored on 2025-05-01 · stale source data (394d)
latest mapped results | top 20
#ModelScoreEvidenceTested
1OpenAI: GPT-5.1
Openai
86.4
model-only
independent_benchmark
2025-05-01
2OpenAI: o3
Openai
85.0
model-only
independent_benchmark
2025-05-01
3Google: Gemini 2.5 Pro
Google
84.5
model-only
independent_benchmark
2025-05-01
4DeepSeek: R1
Deepseek
84.0
model-only
independent_benchmark
2025-05-01
5OpenAI: GPT-4.1
Openai
81.8
model-only
independent_benchmark
2025-05-01
6DeepSeek: DeepSeek V3.2
Deepseek
81.3
model-only
independent_benchmark
2025-05-01
7OpenAI: o1
Openai
80.3
model-only
independent_benchmark
2025-05-01
8OpenAI: o3 Mini
Openai
79.4
model-only
independent_benchmark
2025-05-01
9Google: Gemini 2.0 Flash Lite
Google
76.2
model-only
independent_benchmark
2025-05-01
10Google: Gemini 2.0 Flash
Google
76.2
model-only
independent_benchmark
2025-05-01
11OpenAI: GPT-5
Openai
72.5
model-only
independent_benchmark
2025-05-01
12Google: Gemma 3 27B
Google
67.5
model-only
independent_benchmark
2025-05-01
13Meta: Llama 3.3 70B Instruct
Meta Llama
65.9
model-only
independent_benchmark
2025-05-01
14Mistral Large
Mistralai
65.9
model-only
independent_benchmark
2025-05-01
15Mistral: Mistral Small 4
Mistralai
48.4
model-only
independent_benchmark
2025-05-01
16Mistral: Mistral Nemo
Mistralai
44.8
model-only
independent_benchmark
2025-05-01
what this result means

Massive Multitask Language Understanding — Professional. 14 subjects, 5-shot CoT.

Frontier scores are compressing. Use it as one input, not a complete answer.

A win here is a win on MMLU-Pro. Broad task pages require independent corroboration before naming a general winner.

source record
category: reasoning
metric: accuracy
matched models: 16
latest source date: 2025-05-01
direction: higher is better
inspect upstream source ->