Compare models

Up to six models side by side, on every leaderboard any of them is on. The best in each row is marked. Share the link to share the comparison.

Models (3/6)

Gemini 4 Argon
Claude Opus 5.5
Claude Fable 5.1

Only benchmarks every model is on

Benchmark	Gemini 4 Argon	Claude Opus 5.5	Claude Fable 5.1
The model
Lab	Google	Anthropic	Anthropic
Released	2026-10-13	2026-09-22	2026-09-01
Weights	Proprietary	Proprietary	Proprietary
Context window	1M	1M	1M
Input, per 1M tokens	$2	$4	$10
Output, per 1M tokens	$10	$20	$50
LMArena, people’s votes
LMArena TextMillions of people compare two anonymous models side by side and vote for the better answer	1526#1 of 351	1516#2 of 351	1513#3 of 351
LMArena WebDevPeople ask two anonymous models to build the same web app, try both and vote for the better one	1674#11 of 124	1807#1 of 124	1743#5 of 124
LMArena VisionSide-by-side votes on prompts that include an image	—	1321#2 of 146	1320#4 of 146
LMArena CodingLMArena’s text votes on prompts that ask for code	1539#2 of 346	1545#1 of 346	1531#7 of 346
LMArena MathLMArena’s text votes on math prompts	1538#1 of 339	1518#6 of 339	1532#3 of 339
LMArena Creative writingLMArena’s text votes on stories, poems and other creative writing	1523#2 of 349	1534#1 of 349	1510#3 of 349
LMArena Hard promptsLMArena’s text votes on the hardest prompts	1538#1 of 351	1536#2 of 351	1528#3 of 351
LMArena Instruction followingLMArena’s text votes on prompts with specific instructions to follow	1528#2 of 351	1539#1 of 351	1519#4 of 351
Vals AI, overall
Vals IndexVals AI’s headline score	69.24%#3 of 27	70.18%#2 of 27	—
Vals AI, coding
Terminal-Bench 4.0Frontier-difficulty tasks an agent completes in a real terminal	57.58%#7 of 55	65.15%#1 of 55	58.08%#5 of 55
Vibe Code BenchBuilding complete web apps from a plain-language brief, graded on whether the app works	91.91%#2 of 115	90.29%#5 of 115	90.26%#6 of 115
Code MigrationRewriting working programs in a new language, graded on whether the new program still passes its tests	70.80%#6 of 28	81.01%#2 of 28	—
LiveCodeBenchFresh competitive programming problems, collected after models’ training cut-offs to avoid contamination	—	—	90.52%#1 of 143
Vals AI, finance
Finance AgentThe core tasks of a financial analyst, done as an agent with research tools	67.66%#2 of 31	67.85%#1 of 31	—
Finance IndexVals AI’s composite of finance tasks	64.89%#2 of 31	65.59%#1 of 31	—
Tax Agent BenchAgentic research into US corporate tax questions	44.56%#5 of 76	45.57%#3 of 76	49.19%#2 of 76
Vals AI, law
Legal IndexVals AI’s composite of legal tasks	76.72%#2 of 28	72.33%#7 of 28	—
Legal Research BenchAgentic US legal research	54.81%#4 of 85	50.48%#7 of 85	55.29%#3 of 85
Vals AI, medicine
MedCodeAssigning the right medical billing codes from clinical notes	58.80%#3 of 107	49.80%#19 of 107	53.51%#8 of 107
Vals AI, reasoning
GPQA DiamondGraduate-level questions in biology, physics and chemistry that experts answer well and search does not	—	—	93.43%#9 of 138
MMLU-ProHarder, reasoning-focused multiple choice across 14 subjects, from law to engineering	—	—	92.38%#1 of 138
MMMUCollege-level questions that need both an image and text	—	—	90.64%#1 of 93
Terminal-Bench ScienceExpert-written scientific research workflows, run by an agent in a terminal	44.29%#5 of 42	47.14%#3 of 42	40.00%#6 of 42

Popular comparisons