Claude Opus 4.6 vs GPT-5.2
Wins 4 of 8 areas
Coding · Agents · Reasoning · Facts
Wins 3 of 8 areas
Images and charts · Math · Long documents
Claude Opus 4.6 wins more areas, narrowly.GPT-5.2 is cheaper and better at math.
Scores updated · 65 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.
- ReasoningHard problems that need careful thinking50Claude Opus 4.65 of 6 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own40Claude Opus 4.64 of 4 tests
- CodingWriting and fixing software41Claude Opus 4.64 of 8 tests · 3 ties
- FactsGetting facts right instead of making them up31Claude Opus 4.63 of 4 tests
- MathCompetition and research-level math05GPT-5.25 of 6 tests · 1 tie
- Images and chartsUnderstanding pictures, charts and video12GPT-5.22 of 4 tests · 1 tie
- Long documentsFinding answers in very long texts01GPT-5.21 of 2 tests · 1 tie
- Following instructionsDoing exactly what it is asked11Even1 each
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.
GPT-5.2 costs 48% less for the same work.
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
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+23.5points ahead
- Common-sense trick questionsSimpleBench+21.8points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+18.4points ahead
Where GPT-5.2 pulls ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+22.3points ahead
- Olympiad math problems with short answersIMOAnswerBench+11points ahead
- Extremely hard research-level math problemsFrontierMath Tier 4+4.9points ahead
Every test, side by side
All 65 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingClaude Opus 4.6
- SWE-bench ProClaude Opus 4.6 by 23.553.429.9+23.5
- LMArena · WebDevClaude Opus 4.6 by 130 rating points15461416+130 rating
- SWE-bench MultilingualClaude Opus 4.6 by 11.177.866.7+11.1
- LiveBench · CodingClaude Opus 4.6 by 2.178.276.1+2.1
- LiveBench · Agentic CodingGPT-5.2 by 1.34950.3+1.3
- SWE-bench Verifiedtie80.880tie
- Terminal-Bench Hardtie46.247tie
- SciCodetie51.952.1tie
AgentsClaude Opus 4.6
- BrowseCompClaude Opus 4.6 by 18.28465.8+18.2
- AA ApexAgentsClaude Opus 4.6 by 103323+10
- GDPValClaude Opus 4.6 by 7.655.948.3+7.6
- MCP AtlasClaude Opus 4.6 by 2.162.760.6+2.1
ReasoningClaude Opus 4.6
- SimpleBenchClaude Opus 4.6 by 21.867.645.8+21.8
- ARC-AGI-2Claude Opus 4.6 by 15.968.852.9+15.9
- LiveBench · ReasoningClaude Opus 4.6 by 5.588.783.2+5.5
- Humanity's Last ExamClaude Opus 4.6 by 2.239.937.7+2.2
- CritPtClaude Opus 4.6 by 112.611.6+1
- GPQA Diamondtie89.690.3tie
FactsClaude Opus 4.6
- AA-Omniscience · Non-hallucinationClaude Opus 4.6 by 18.437.218.8+18.4
- SimpleQA VerifiedClaude Opus 4.6 by 9.94737.1+9.9
- AA-Omniscience · AccuracyClaude Opus 4.6 by 2.74744.3+2.7
- Vectara HHEM hallucination ratelower is betterGPT-5.2 by 1.412.210.8+1.4
Images and chartsGPT-5.2
- LMArena · VisionClaude Opus 4.6 by 70 rating points13161246+70 rating
- CharXiv (RQ)GPT-5.2 by 4.777.482.1+4.7
- OCRBenchv2GPT-5.2 by 2.148.450.5+2.1
- MMMU-Protie75.474.6tie
MathGPT-5.2
- IMOAnswerBenchGPT-5.2 by 1175.386.3+11
- FrontierMath Tier 4GPT-5.2 by 4.926.831.7+4.9
- LiveBench · MathematicsGPT-5.2 by 3.989.393.2+3.9
- AIME 2026GPT-5.2 by 1.696.798.3+1.6
- FrontierMath Tiers 1-3 (v2)GPT-5.2 by 1.46667.4+1.4
- HMMT Feb. 2026tie96.297tie
Long documentsGPT-5.2
- AA-LCRGPT-5.2 by 4.77882.7+4.7
- MRCR v2 8 needle 128k (average)tie8483.8tie
Following instructionsEven
- IFBenchGPT-5.2 by 22.353.175.4+22.3
- LiveBench · Instruction FollowingClaude Opus 4.6 by 1.563.361.8+1.5
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.
- ScreenSpot-Pro (No tools)GPT-5.2 by 28.657.786.3+28.6
- DeepSearchQA (F1)Claude Opus 4.6 by 2091.371.3+20
- MATH-VisionGPT-5.2 by 17.565.583+17.5
- GDPval-AA (Elo)Claude Opus 4.6 by 157 rating points16191462+157 rating
- AA-OmniscienceClaude Opus 4.6 by 14.613.7-0.9+14.6
- matharena_visual_math_overallGPT-5.2 by 14.272.386.5+14.2
- Terminal-Bench 2.0Claude Opus 4.6 by 11.465.454+11.4
- ToolathlonClaude Opus 4.6 by 10.556.846.3+10.5
- τ²-Bench (Retail)Claude Opus 4.6 by 9.991.982+9.9
- livebench_data_analysisGPT-5.2 by 8.369.978.2+8.3
- HLE (with tools)Claude Opus 4.6 by 7.853.345.5+7.8
- AA Agentic IndexClaude Opus 4.6 by 7.467.660.2+7.4
- τ²-Bench Telecom (AA run)Claude Opus 4.6 by 7.392.184.8+7.3
- ARC-AGI-1Claude Opus 4.6 by 6.89386.2+6.8
- vectara_avg_summary_lengthGPT-5.2 by 48.7 rating points137.6186.3+48.7 rating
- frontiermath_tier_4_v1Claude Opus 4.6 by 4.122.918.8+4.1
- livebench_languageClaude Opus 4.6 by 3.583.379.8+3.5
- AIME25 no toolsGPT-5.2 by 2.495.698+2.4
- MMLU-ProClaude Opus 4.6 by 2.489.186.7+2.4
- AA IntelligenceClaude Opus 4.6 by 1.531.930.4+1.5
- LiveBenchClaude Opus 4.6 by 1.576.374.8+1.5
- MMMLUClaude Opus 4.6 by 1.591.189.6+1.5
- vectara_factual_consistencyGPT-5.2 by 1.487.889.2+1.4
- metr_swaaGPT-5.2 by 1.198.599.6+1.1
- Artificial Analysis Coding Indextie48.148.7tie
- HMMT Nov. 2025tie96.395.8tie
- AIME 2025tie99.8100tie
- LiveCodeBenchtie88.889tie
- vectara_answer_ratetie99.8100tie
Questions people ask
Which is better, Claude Opus 4.6 or GPT-5.2?
Claude Opus 4.6 wins four of the eight areas we test: coding, agents, reasoning and facts. GPT-5.2 wins images and charts, math and long documents, and costs 48% less. They are level on following instructions.
Which is better for coding?
Claude Opus 4.6. It wins 4 of the 8 coding tests both models report; GPT-5.2 wins 1, and 3 are ties.
Which is cheaper?
Claude Opus 4.6 costs $5.00 per million input tokens and $25.00 per million output tokens; GPT-5.2 costs $1.75 and $14.00. That makes GPT-5.2 about 48% cheaper for the same work.
How do you compare the two?
We use the 65 benchmark tests both models have published scores on. The verdict counts the 36 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.