Claude Opus 4.5 vs Claude Opus 4.6
Wins 2 of 8 areas
Facts · Following instructions
Wins 5 of 8 areas
Coding · Agents · Reasoning · Images and charts · Math
Claude Opus 4.6 is the stronger all-rounder.Claude Opus 4.5 is better at following instructions.
Scores updated · 63 tests both models report · How we compare
Where each one wins
Tests won in each of the eight areas we test. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- CodingWriting and fixing software16Claude Opus 4.66 of 9 tests · 2 ties
- ReasoningHard problems that need careful thinking05Claude Opus 4.65 of 5 tests
- AgentsCarrying out multi-step tasks on its own03Claude Opus 4.63 of 4 tests · 1 tie
- MathCompetition and research-level math24Claude Opus 4.64 of 6 tests
- Images and chartsUnderstanding pictures, charts and video02Claude Opus 4.62 of 2 tests
- Following instructionsDoing exactly what it is asked20Claude Opus 4.52 of 3 tests · 1 tie
- FactsGetting facts right instead of making them up21Claude Opus 4.52 of 4 tests · 1 tie
- Long documentsFinding answers in very long texts00Even0 each · 1 tie
Long documents rests on a single test.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Claude Opus 4.6 pulls ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+70.8points ahead
- Finds hard-to-locate facts by browsing the webBrowseComp+47points ahead
- Unpublished advanced math problemsFrontierMath Tiers 1-3 (v2)+31.6points ahead
Where Claude Opus 4.5 pulls ahead
- Keeps track of context across a multi-turn chatMulti-Challenge+21.8points ahead
- Olympiad math problems with short answersIMOAnswerBench+8.7points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+4.9points ahead
Every test, side by side
All 63 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingClaude Opus 4.6
- SWE-bench MultilingualClaude Opus 4.6 by 70.8777.8+70.8
- LiveBench · Agentic CodingClaude Opus 4.6 by 9.339.749+9.3
- LMArena · WebDevClaude Opus 4.6 by 77 rating points14691546+77 rating
- LiveCodeBench v6Claude Opus 4.6 by 484.888.8+4
- SciCodeClaude Opus 4.6 by 2.449.551.9+2.4
- LiveBench · CodingClaude Opus 4.5 by 1.579.778.2+1.5
- SWE-bench ProClaude Opus 4.6 by 1.45253.4+1.4
- Terminal-Bench Hardtie4746.2tie
- SWE-bench Verifiedtie80.980.8tie
AgentsClaude Opus 4.6
- BrowseCompClaude Opus 4.6 by 473784+47
- GDPValClaude Opus 4.6 by 8.647.355.9+8.6
- OSWorld-VerifiedClaude Opus 4.6 by 6.466.372.7+6.4
- MCP Atlastie62.362.7tie
ReasoningClaude Opus 4.6
- ARC-AGI-2Claude Opus 4.6 by 31.237.668.8+31.2
- Humanity's Last ExamClaude Opus 4.6 by 9.830.139.9+9.8
- LiveBench · ReasoningClaude Opus 4.6 by 8.680.188.7+8.6
- CritPtClaude Opus 4.6 by 84.612.6+8
- GPQA DiamondClaude Opus 4.6 by 386.689.6+3
FactsClaude Opus 4.5
- AA-Omniscience · Non-hallucinationClaude Opus 4.5 by 1.83937.2+1.8
- SimpleQA VerifiedClaude Opus 4.6 by 1.345.747+1.3
- Vectara HHEM hallucination ratelower is betterClaude Opus 4.5 by 1.310.912.2+1.3
- AA-Omniscience · Accuracytie46.647tie
Images and chartsClaude Opus 4.6
- CharXiv (RQ)Claude Opus 4.6 by 10.267.277.4+10.2
- MMMU-ProClaude Opus 4.6 by 1.47475.4+1.4
MathClaude Opus 4.6
- FrontierMath Tiers 1-3 (v2)Claude Opus 4.6 by 31.634.466+31.6
- FrontierMath Tier 4Claude Opus 4.6 by 21.94.926.8+21.9
- HMMT Feb. 2026Claude Opus 4.6 by 10.985.396.2+10.9
- IMOAnswerBenchClaude Opus 4.5 by 8.78475.3+8.7
- AIME 2026Claude Opus 4.6 by 1.695.196.7+1.6
- LiveBench · MathematicsClaude Opus 4.5 by 1.190.489.3+1.1
Following instructionsClaude Opus 4.5
- Multi-ChallengeClaude Opus 4.5 by 21.85937.1+21.8
- IFBenchClaude Opus 4.5 by 4.95853.1+4.9
- LiveBench · Instruction Followingtie62.563.3tie
Other results29 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- CyberGymClaude Opus 4.6 by 23.250.673.8+23.2
- frontiermath_tier_4_v1Claude Opus 4.6 by 18.74.222.9+18.7
- DeepSearchQA (F1)Claude Opus 4.6 by 15.276.191.3+15.2
- ToolathlonClaude Opus 4.6 by 13.343.556.8+13.3
- MATH-VisionClaude Opus 4.5 by 11.677.165.5+11.6
- ARC-AGI-1Claude Opus 4.6 by 118293+11
- HLE (with tools)Claude Opus 4.6 by 10.143.253.3+10.1
- AA Agentic IndexClaude Opus 4.6 by 859.667.6+8
- AIME 2025Claude Opus 4.6 by 792.899.8+7
- NL2RepoClaude Opus 4.6 by 6.643.249.8+6.6
- Terminal-Bench 2.0Claude Opus 4.6 by 6.159.365.4+6.1
- Finance AgentClaude Opus 4.6 by 5.555.260.7+5.5
- AIME25 no toolsClaude Opus 4.6 by 4.69195.6+4.6
- livebench_data_analysisClaude Opus 4.5 by 4.574.469.9+4.5
- HMMT Nov. 2025Claude Opus 4.6 by 393.396.3+3
- τ²-Bench (Retail)Claude Opus 4.6 by 388.991.9+3
- AA IntelligenceClaude Opus 4.6 by 2.829.131.9+2.8
- τ²-Bench Telecom (AA run)Claude Opus 4.6 by 2.689.592.1+2.6
- vectara_avg_summary_lengthClaude Opus 4.6 by 23.1 rating points114.5137.6+23.1 rating
- livebench_languageClaude Opus 4.6 by 281.383.3+2
- LiveCodeBenchClaude Opus 4.6 by 1.88788.8+1.8
- vectara_factual_consistencyClaude Opus 4.5 by 1.389.187.8+1.3
- vectara_answer_rateClaude Opus 4.6 by 1.198.799.8+1.1
- metr_swaaClaude Opus 4.5 by 199.598.5+1
- MMLU-Protie89.589.1tie
- AA-Omnisciencetie1413.7tie
- Artificial Analysis Coding Indextie47.848.1tie
- LiveBenchtie7676.3tie
- MMMLUtie90.891.1tie
Questions people ask
Which is better, Claude Opus 4.5 or Claude Opus 4.6?
Claude Opus 4.6 wins five of the eight areas we test: coding, agents, reasoning, images and charts and math. Claude Opus 4.5 wins facts and following instructions. They are level on long documents.
Which is better for coding?
Claude Opus 4.6. It wins 6 of the 9 coding tests both models report; Claude Opus 4.5 wins 1, and 2 are ties.
Which is cheaper?
Claude Opus 4.5 costs $5.00 per million input tokens and $25.00 per million output tokens; Claude Opus 4.6 costs $5.00 and $25.00. For a million tokens read plus a million written, they cost about the same.
How do you compare the two?
We use the 63 benchmark tests both models have published scores on. The verdict counts the 34 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 29 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.