Benchmark Data — April 2026

LLM Benchmark Scores 2026 — MMLU, HumanEval, MATH & More

Standardized benchmark results for 16 leading AI models across 8 widely recognised evaluation suites. Scores are sourced from official model cards, research papers, and reproducible third-party evaluations as of April 2026.

MMLU
57 academic subjects from elementary to graduate level.
Scale: 0–100%
HumanEval
164 hand-written Python programming problems.
Scale: 0–100%
MATH
12,500 competition math problems across 5 difficulty levels.
Scale: 0–100%
GPQA
448 expert-level questions in biology, chemistry, and physics.
Scale: 0–100%
GSM8K
8,500 linguistically diverse grade-school math problems.
Scale: 0–100%
SWE-bench
300 real software engineering issues from open-source repos.
Scale: 0–100%
MT-Bench
80 multi-turn questions judged by GPT-4 on a 1–10 scale.
Scale: 1–10
MMMU
11,500 university-level multimodal questions across 30 subjects.
Scale: 0–100%

All Benchmark Results

Model Provider MMLU HumanEval MATH GPQA GSM8K SWE-bench MT-Bench MMMU
Claude Opus 4.6 Anthropic 92.1 92.4 91.5 74.2 97.8 58.4 9.4 82.1
GPT-5.4 OpenAI 91.8 94.1 90.2 72.8 98.1 62.3 9.5 80.4
Gemini 3.1 Ultra Google 90.4 89.3 88.7 71.5 96.4 54.2 9.3 86.3
o3-pro OpenAI 88.2 91.2 96.7 87.4 99.1 60.1 9.1 76.2
Claude Sonnet 4.6 Anthropic 89.7 90.8 86.4 68.3 95.6 54.7 9.2 79.3
Gemini 3.1 Pro Google 88.1 87.6 85.3 67.1 94.2 51.3 9.0 84.1
GPT-5.4 mini OpenAI 85.3 88.2 80.1 60.2 91.4 46.8 8.7 74.1
Grok 3.5 xAI 86.4 84.3 82.7 63.4 92.8 44.2 8.8 72.4
Llama 4 Maverick Meta 84.7 82.1 78.4 58.3 90.1 42.6 8.5 76.8
DeepSeek V4 DeepSeek 87.2 88.7 84.1 65.2 93.7 48.3 8.9 70.2
Claude Haiku 4.5 Anthropic 80.4 81.3 74.2 52.1 86.4 38.7 8.3 72.1
Gemini 3.1 Flash Google 79.8 78.4 72.1 50.3 84.7 36.4 8.1 78.4
Qwen3 72B Alibaba 83.6 85.4 80.7 61.2 90.3 44.7 8.6 68.3
Mistral Large 2 Mistral 78.2 76.8 68.4 46.8 82.1 34.2 7.9 62.4
Phi-4 Microsoft 76.8 78.2 72.8 44.2 84.3 36.8 8.0 60.1
Llama 3.3 70B Meta 75.4 74.6 66.2 42.1 80.8 32.4 7.7 60.8
High score Mid score Lower score Best in benchmark

Top Performers per Benchmark

MMLU
Anthropic
92.1 0–100%
HumanEval
OpenAI
94.1 0–100%
MATH
OpenAI
96.7 0–100%
GPQA
OpenAI
87.4 0–100%
GSM8K
OpenAI
99.1 0–100%
SWE-bench
OpenAI
62.3 0–100%
MT-Bench
OpenAI
9.5 1–10
MMMU
86.3 0–100%

About Each Benchmark

MMLU — Massive Multitask Language Understanding
57 academic subjects from elementary to graduate level. Scale: 0–100%

Tests breadth of knowledge. High scores indicate strong general academic understanding across STEM, humanities, and professional domains.

#1
Claude Opus 4.6 92.1
#2
GPT-5.4 91.8
#3
Gemini 3.1 Ultra 90.4
HumanEval — Code Generation from Docstrings
164 hand-written Python programming problems. Scale: 0–100%

Evaluates ability to write correct code from docstrings. Directly relevant for coding assistants and software development use cases.

#1
GPT-5.4 94.1
#2
Claude Opus 4.6 92.4
#3
o3-pro 91.2
MATH — Competition Mathematics
12,500 competition math problems across 5 difficulty levels. Scale: 0–100%

Measures mathematical reasoning from AMC to AIME difficulty. Correlates with broader reasoning capability.

#1
o3-pro 96.7
#2
Claude Opus 4.6 91.5
#3
GPT-5.4 90.2
GPQA — Graduate-Level Science Q&A
448 expert-level questions in biology, chemistry, and physics. Scale: 0–100%

Expert-level science questions that stump even PhD students. A tough frontier test of genuine understanding vs. pattern matching.

#1
o3-pro 87.4
#2
Claude Opus 4.6 74.2
#3
GPT-5.4 72.8
GSM8K — Grade School Math Word Problems
8,500 linguistically diverse grade-school math problems. Scale: 0–100%

Tests multi-step arithmetic reasoning in plain language. Near-saturated for frontier models — now mainly a baseline check.

#1
o3-pro 99.1
#2
GPT-5.4 98.1
#3
Claude Opus 4.6 97.8
SWE-bench — Real GitHub Issue Resolution
300 real software engineering issues from open-source repos. Scale: 0–100%

Measures real-world software engineering. The gold standard for agentic coding: resolve GitHub issues end-to-end.

#1
GPT-5.4 62.3
#2
o3-pro 60.1
#3
Claude Opus 4.6 58.4
MT-Bench — Multi-Turn Instruction Following
80 multi-turn questions judged by GPT-4 on a 1–10 scale. Scale: 1–10

LLM-judged multi-turn conversations. Reflects instruction-following quality in realistic chat and assistant scenarios.

#1
GPT-5.4 9.5
#2
Claude Opus 4.6 9.4
#3
Gemini 3.1 Ultra 9.3
MMMU — Multimodal University Questions
11,500 university-level multimodal questions across 30 subjects. Scale: 0–100%

The leading multimodal benchmark. Tests chart understanding, scientific diagrams, and cross-modal reasoning simultaneously.

#1
Gemini 3.1 Ultra 86.3
#2
Gemini 3.1 Pro 84.1
#3
Claude Opus 4.6 82.1