Compare models. Find the differences.

Add, remove, or change models in the table.

Benchmark snapshot

1 benchmarks in common.

4 models selected. 1 benchmarks in common.

Reported scores; test settings may differ.

Reported benchmark results for v3 V3.2 Exp Chat, 3.5 27B Thinking, Claude Haiku 5.5 Thinking, M2 M2.7
Benchmark
DeepSeek
Qwen
Anthropic
MiniMax
Arena ELOHuman preference1,4221,4091,4251,415
LiveBench 1.0General reasoningNot in our datasetNot in our dataset72.1%Not in our dataset
GPQA DiamondScientific knowledgeNot in our dataset85.5%Not in our dataset87.0%
HLEAcademic reasoningNot in our dataset24.3%30.1%28.0%
MMLU ProMulti-task understandingNot in our dataset86.1%Not in our dataset81.9%
AiderCode editing70.2%Not in our datasetNot in our datasetNot in our dataset
SWE-Bench VerifiedAgentic codingNot in our dataset75.0%Not in our dataset72.2%
Tau-BenchAgentic tool useNot in our datasetNot in our datasetNot in our dataset84.9%
ARC-AGI-2Visual reasoningNot in our datasetNot in our dataset61.3%Not in our dataset
Pricing & context Published USD rates per 1M tokens
Specificationv3 V3.2 Exp Chat3.5 27B ThinkingClaude Haiku 5.5 ThinkingM2 M2.7
Input$0.27$0.26$0.1$0.255
Output$0.41$2.6$0.5$1.02
Context window164K262K1.0M205K
Pricing detailsExplore pricing →Explore pricing →Explore pricing →Explore pricing →
How to read this comparison

Each column keeps the exact model version you selected. Shared benchmarks are the intersection of their reported results; “all coverage” also shows where evidence is missing. Missing results never count as zero.

These are published benchmark scores, not a controlled head-to-head test. The available evidence does not establish matching evaluation settings, so we show no cross-model winner highlights or deltas. Composite quality scores are omitted because they can summarize different benchmark sets.

Scores and statistics are computed upstream. This page reads a fixed data snapshot; its date describes the snapshot, not the date of every measurement.