Model Benchmarks

How the models score on public capability benchmarks — and, for each, the features it measures and the real-world use cases those features unlock. Click a benchmark for the full leaderboard.

Model Library View model pricing View benchmarks

Who leads across the benchmarks

#1 finishes by model family across all 18 benchmarks. Boards rarely share the same models, so this counts each benchmark's top model rather than a single cross-board score.

Claude · Anthropic 10 / 18
Gemini · Google 3 / 18
Open / research models 2 / 18
GPT / o-series · OpenAI 2 / 18
Mistral 1 / 18

Hover a family to see which boards it tops. The full per-benchmark leaderboards are below.

Filter each column from its heading. Type to search — Category also offers a dropdown of values.

Features measured Models
MMLU Knowledge & Reasoning Reasoning Multilingual 5 Claude Opus 5 89.2%
GPQA Diamond Knowledge & Reasoning Reasoning 4 Claude Opus 5 62.4%
SWE-bench Verified Coding & Agents Code Tool use Reasoning Long context 4 Claude Opus 5 74.5%
HumanEval Coding & Agents Code 4 Claude Opus 5 94.1%
MMMU Multimodal Vision Reasoning 3 Gemini 2.5 Pro 74.3%
AIME 2025 Math Reasoning 3 Claude Opus 5 80%
MTEB Retrieval Embeddings 1 E5 Mistral 7B (embeddings) 66.6
MGSM Multilingual Multilingual Reasoning 3 Gemini 2.5 Pro 91.5%
LMArena (Text Arena) Knowledge & Reasoning Reasoning Multilingual 11 Claude Mythos 5 1,531
Terminal-Bench 2.1 Coding & Agents Tool use Code Reasoning Long context 8 Claude Code + Fable 5 83.8%
τ²-bench (Telecom) Coding & Agents Tool use Reasoning 4 Claude 3.7 Sonnet 49%
Video-MME Multimodal Vision Reasoning 5 video-SALMONN 2+ 79.7%
SciCode Coding & Agents Code Reasoning 6 OpenAI o3-mini (low) 10.8%
CharXiv (Reasoning) Multimodal Vision Reasoning 7 Claude 3.5 Sonnet 60.2%
MMMLU (Multilingual) Multilingual Multilingual Reasoning 8 o3 (high) 88.8%
GDPval Knowledge & Reasoning Reasoning Tool use 1 Claude Opus 4.1 47.6%
SWE-bench Multilingual Coding & Agents Code Tool use Reasoning Long context Multilingual 8 Gemini 3 Flash 72.7%
OmniDocBench Multimodal Vision 8 PaddleOCR-VL-1.6 96.3

One Gateway, Every Model

Get an API key and point your first request at ApiSpi in minutes