GPT-5.2 vs GPT-5.4
Wins 0 of 8 areas
—
Wins 6 of 8 areas
Coding · Agents · Reasoning · Facts · Images and charts · Math
GPT-5.4 is the stronger all-rounder.GPT-5.2 is cheaper.
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.
- CodingWriting and fixing software07GPT-5.47 of 7 tests
- ReasoningHard problems that need careful thinking05GPT-5.45 of 5 tests
- MathCompetition and research-level math04GPT-5.44 of 6 tests · 2 ties
- AgentsCarrying out multi-step tasks on its own14GPT-5.44 of 5 tests
- FactsGetting facts right instead of making them up13GPT-5.43 of 4 tests
- Images and chartsUnderstanding pictures, charts and video02GPT-5.42 of 3 tests · 1 tie
- Long documentsFinding answers in very long texts00Even0 each · 1 tie
- Following instructionsDoing exactly what it is asked11Even1 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-5.2 costs 10% 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
- Customer-service tasks in a simulated bankτ-Bench V3 · Banking+28.5points ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+27.8points ahead
- Abstract visual puzzles that people can solveARC-AGI-2+20.4points ahead
Where GPT-5.2 pulls ahead
- Real work tasks from 44 professionsGDPVal+10.9points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+10.5points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+1.5points 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.
CodingGPT-5.4
- SWE-bench ProGPT-5.4 by 27.829.957.7+27.8
- Terminal-Bench HardGPT-5.4 by 10.64757.6+10.6
- SWE-bench MultilingualGPT-5.4 by 566.771.7+5
- LMArena · WebDevGPT-5.4 by 49 rating points14161465+49 rating
- SciCodeGPT-5.4 by 4.552.156.6+4.5
- LiveBench · Agentic CodingGPT-5.4 by 3.550.353.8+3.5
- LiveBench · CodingGPT-5.4 by 1.476.177.5+1.4
AgentsGPT-5.4
- τ-Bench V3 · BankingGPT-5.4 by 28.511.139.6+28.5
- BrowseCompGPT-5.4 by 16.965.882.7+16.9
- GDPValGPT-5.2 by 10.948.337.4+10.9
- AA ApexAgentsGPT-5.4 by 10.32333.3+10.3
- MCP AtlasGPT-5.4 by 6.660.667.2+6.6
ReasoningGPT-5.4
- ARC-AGI-2GPT-5.4 by 20.452.973.3+20.4
- CritPtGPT-5.4 by 11.811.623.4+11.8
- Humanity's Last ExamGPT-5.4 by 637.743.7+6
- LiveBench · ReasoningGPT-5.4 by 4.983.288.1+4.9
- GPQA DiamondGPT-5.4 by 1.790.392+1.7
FactsGPT-5.4
- AA-Omniscience · Non-hallucinationGPT-5.2 by 10.518.88.3+10.5
- SimpleQA VerifiedGPT-5.4 by 837.145.1+8
- AA-Omniscience · AccuracyGPT-5.4 by 6.644.350.9+6.6
- Vectara HHEM hallucination ratelower is betterGPT-5.4 by 3.810.87+3.8
Images and chartsGPT-5.4
- LMArena · VisionGPT-5.4 by 56 rating points12461302+56 rating
- MMMU-ProGPT-5.4 by 3.874.678.4+3.8
- CharXiv (RQ)tie82.182.8tie
MathGPT-5.4
- FrontierMath Tier 4GPT-5.4 by 17.331.749+17.3
- FrontierMath Tiers 1-3 (v2)GPT-5.4 by 11.267.478.6+11.2
- IMOAnswerBenchGPT-5.4 by 5.186.391.4+5.1
- LiveBench · MathematicsGPT-5.4 by 193.294.2+1
- AIME 2026tie98.399.2tie
- HMMT Feb. 2026tie9797.7tie
Following instructionsEven
- LiveBench · Instruction FollowingGPT-5.4 by 8.461.870.2+8.4
- IFBenchGPT-5.2 by 1.575.473.9+1.5
Other results32 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.
- Artificial Analysis Coding IndexGPT-5.4 by 22.448.771.1+22.4
- GDPval-AA (Elo)GPT-5.4 by 212 rating points14621674+212 rating
- Terminal-Bench 2.0GPT-5.4 by 21.15475.1+21.1
- AA Agentic IndexGPT-5.2 by 1660.244.2+16
- vectara_avg_summary_lengthGPT-5.2 by 104.6 rating points186.381.7+104.6 rating
- MATH-VisionGPT-5.4 by 98392+9
- AA IntelligenceGPT-5.4 by 8.630.439+8.6
- frontiermath_tier_4_v1GPT-5.4 by 8.318.827.1+8.3
- ToolathlonGPT-5.4 by 8.346.354.6+8.3
- ARC-AGI-1GPT-5.4 by 7.586.293.7+7.5
- DeepSearchQA (F1)GPT-5.4 by 7.371.378.6+7.3
- FrontierMath (overall)GPT-5.4 by 7.340.347.6+7.3
- AA-OmniscienceGPT-5.4 by 6.7-0.95.8+6.7
- HLE (with tools)GPT-5.4 by 6.645.552.1+6.6
- matharena_visual_math_overallGPT-5.4 by 686.592.5+6
- LiveBenchGPT-5.4 by 5.574.880.3+5.5
- HealthBench Hard length-adjustedGPT-5.2 by 5.234.329.1+5.2
- vectara_factual_consistencyGPT-5.4 by 3.889.293+3.8
- OmniDocBench 1.5GPT-5.4 by 3.485.789.1+3.4
- HealthBench length-adjustedGPT-5.2 by 2.856.854+2.8
- livebench_languageGPT-5.4 by 2.879.882.6+2.8
- τ²-Bench Telecom (AA run)GPT-5.4 by 2.384.887.1+2.3
- HealthBench Professional length-adjustedGPT-5.4 by 2.245.948.1+2.2
- HealthBench Consensus length-adjustedGPT-5.4 by 1.994.496.3+1.9
- livebench_data_analysisGPT-5.4 by 1.178.279.3+1.1
- MMLU-Protie86.787.5tie
- vectara_answer_ratetie10099.9tie
- HMMT Nov. 2025tie95.895.8tie
- Image input eval - extremism (not_unsafe)tie11tie
- Image input eval - harms-erotic (not_unsafe)tie11tie
- Image input eval - hate (not_unsafe)tie11tie
- Image input eval - self-harm (not_unsafe)tie11tie
Questions people ask
Which is better, GPT-5.2 or GPT-5.4?
GPT-5.4 wins six of the eight areas we test: coding, agents, reasoning, facts, images and charts and math. GPT-5.2 wins none, but costs 10% less. They are level on long documents and following instructions.
Which is better for coding?
GPT-5.4. It wins 7 of the 7 coding tests both models report; GPT-5.2 wins none.
Which is cheaper?
GPT-5.2 costs $1.75 per million input tokens and $14.00 per million output tokens; GPT-5.4 costs $2.50 and $15.00. That makes GPT-5.2 about 10% 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 33 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 32 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.