Claude Opus 4.7 vs GPT-5.4
Wins 3 of 8 areas
Coding · Agents · Images and charts
Wins 4 of 8 areas
Reasoning · Math · Long documents · Following instructions
GPT-5.4 wins more areas, narrowly.Claude Opus 4.7 is better at coding.
Scores updated · 61 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.
- MathCompetition and research-level math06GPT-5.46 of 6 tests
- Following instructionsDoing exactly what it is asked02GPT-5.42 of 2 tests
- ReasoningHard problems that need careful thinking12GPT-5.42 of 6 tests · 3 ties
- Long documentsFinding answers in very long texts01GPT-5.41 of 1 test
- CodingWriting and fixing software53Claude Opus 4.75 of 8 tests
- AgentsCarrying out multi-step tasks on its own42Claude Opus 4.74 of 7 tests · 1 tie
- Images and chartsUnderstanding pictures, charts and video20Claude Opus 4.72 of 3 tests · 1 tie
- FactsGetting facts right instead of making them up22Even2 each
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.
GPT-5.4 costs 42% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where GPT-5.4 pulls ahead
- Proof problems from the US Math OlympiadUSAMO 2026+25.9points ahead
- Extremely hard research-level math problemsFrontierMath Tier 4+17.3points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+15.3points ahead
Where Claude Opus 4.7 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+49.4points ahead
- Finds the root cause of IT system incidentsAA IT-Bench SRE+12.2points ahead
- Multi-step tasks using many real software toolsMCP Atlas+10.1points ahead
Every test, side by side
All 61 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingClaude Opus 4.7
- LMArena · WebDevClaude Opus 4.7 by 92 rating points15571465+92 rating
- SWE-bench MultilingualClaude Opus 4.7 by 8.880.571.7+8.8
- SWE-bench ProClaude Opus 4.7 by 6.664.357.7+6.6
- Terminal-Bench HardGPT-5.4 by 6.151.557.6+6.1
- Terminal-Bench 2.1Claude Opus 4.7 by 4.883.178.3+4.8
- LiveBench · CodingClaude Opus 4.7 by 4.682.177.5+4.6
- LiveBench · Agentic CodingGPT-5.4 by 3.150.753.8+3.1
- SciCodeGPT-5.4 by 2.154.556.6+2.1
AgentsClaude Opus 4.7
- AA IT-Bench SREClaude Opus 4.7 by 12.246.734.5+12.2
- MCP AtlasClaude Opus 4.7 by 10.177.367.2+10.1
- GDPValClaude Opus 4.7 by 5.442.837.4+5.4
- τ-Bench V3 · BankingGPT-5.4 by 534.639.6+5
- BrowseCompGPT-5.4 by 3.479.382.7+3.4
- OSWorld-VerifiedClaude Opus 4.7 by 37875+3
- AA ApexAgentstie33.933.3tie
ReasoningGPT-5.4
- CritPtGPT-5.4 by 11.41223.4+11.4
- ARC-AGI-2Claude Opus 4.7 by 2.575.873.3+2.5
- Humanity's Last ExamGPT-5.4 by 1.442.343.7+1.4
- LiveBench · Reasoningtie87.288.1tie
- GPQA Diamondtie91.492tie
- ARC-AGI-3tie0.20.2tie
FactsEven
- AA-Omniscience · Non-hallucinationClaude Opus 4.7 by 49.457.78.3+49.4
- SimpleQA VerifiedClaude Opus 4.7 by 6.651.745.1+6.6
- Vectara HHEM hallucination ratelower is betterGPT-5.4 by 5127+5
- AA-Omniscience · AccuracyGPT-5.4 by 248.950.9+2
Images and chartsClaude Opus 4.7
- CharXiv (RQ)Claude Opus 4.7 by 8.29182.8+8.2
- LMArena · VisionClaude Opus 4.7 by 11 rating points13131302+11 rating
- MMMU-Protie78.878.4tie
MathGPT-5.4
- USAMO 2026GPT-5.4 by 25.969.395.2+25.9
- FrontierMath Tier 4GPT-5.4 by 17.331.749+17.3
- FrontierMath Tiers 1-3 (v2)GPT-5.4 by 8.470.278.6+8.4
- HMMT Feb. 2026GPT-5.4 by 3.893.997.7+3.8
- AIME 2026GPT-5.4 by 3.495.899.2+3.4
- LiveBench · MathematicsGPT-5.4 by 1.392.894.2+1.3
Following instructionsGPT-5.4
- IFBenchGPT-5.4 by 15.358.673.9+15.3
- LiveBench · Instruction FollowingGPT-5.4 by 3.566.770.2+3.5
Other results24 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.
- HiL-Bench (Tools-allowed)Claude Opus 4.7 by 3241.79.7+32
- OfficeQA ProClaude Opus 4.7 by 29.580.651.1+29.5
- BabyVisionGPT-5.4 by 27.522.249.7+27.5
- AA-OmniscienceClaude Opus 4.7 by 21.527.35.8+21.5
- OfficeQAClaude Opus 4.7 by 18.286.368.1+18.2
- Finance AgentClaude Opus 4.7 by 8.464.456+8.4
- GDPval-AA (Elo)Claude Opus 4.7 by 79 rating points17531674+79 rating
- CyberGymClaude Opus 4.7 by 6.873.166.3+6.8
- vectara_avg_summary_lengthClaude Opus 4.7 by 67.4 rating points149.181.7+67.4 rating
- Legal Agent BenchmarkClaude Opus 4.7 by 6.77.10.4+6.7
- Terminal-Bench 2.0GPT-5.4 by 5.769.475.1+5.7
- vectara_factual_consistencyGPT-5.4 by 58893+5
- AA Agentic IndexGPT-5.4 by 4.739.544.2+4.7
- livebench_languageGPT-5.4 by 4.777.982.6+4.7
- ToolathlonClaude Opus 4.7 by 4.759.354.6+4.7
- frontiermath_tier_4_v1GPT-5.4 by 4.222.927.1+4.2
- LiveBenchGPT-5.4 by 3.476.980.3+3.4
- HLE (with tools)Claude Opus 4.7 by 2.754.752+2.7
- Artificial Analysis Coding IndexClaude Opus 4.7 by 2.573.671.1+2.5
- vectara_answer_rateGPT-5.4 by 1.99899.9+1.9
- AA IntelligenceClaude Opus 4.7 by 1.740.739+1.7
- ARC-AGI-1GPT-5.4 by 1.79293.7+1.7
- τ²-Bench Telecom (AA run)Claude Opus 4.7 by 1.588.687.1+1.5
- livebench_data_analysisGPT-5.4 by 178.379.3+1
Questions people ask
Which is better, Claude Opus 4.7 or GPT-5.4?
GPT-5.4 wins four of the eight areas we test: reasoning, math, long documents and following instructions. Claude Opus 4.7 wins coding, agents and images and charts. They are level on facts.
Which is better for coding?
Claude Opus 4.7. It wins 5 of the 8 coding tests both models report; GPT-5.4 wins 3.
Which is cheaper?
Claude Opus 4.7 costs $5.00 per million input tokens and $25.00 per million output tokens; GPT-5.4 costs $2.50 and $15.00. That makes GPT-5.4 about 42% cheaper for the same work.
How do you compare the two?
We use the 61 benchmark tests both models have published scores on. The verdict counts the 37 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 24 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.