GPT-5.4 vs GPT-5.5
Wins 0 of 8 areas
—
Wins 7 of 8 areas
Coding · Agents · Reasoning · Facts · Math · Long documents · Following instructions
GPT-5.5 is the stronger all-rounder.GPT-5.4 is cheaper.
Scores updated · 71 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.
- AgentsCarrying out multi-step tasks on its own06GPT-5.56 of 7 tests · 1 tie
- ReasoningHard problems that need careful thinking05GPT-5.55 of 6 tests · 1 tie
- CodingWriting and fixing software04GPT-5.54 of 7 tests · 3 ties
- MathCompetition and research-level math04GPT-5.54 of 6 tests · 2 ties
- FactsGetting facts right instead of making them up03GPT-5.53 of 3 tests
- Following instructionsDoing exactly what it is asked01GPT-5.51 of 2 tests · 1 tie
- Long documentsFinding answers in very long texts01GPT-5.51 of 1 test
- Images and chartsUnderstanding pictures, charts and video11Even1 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 50% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where GPT-5.5 pulls ahead
- Extremely hard research-level math problemsFrontierMath Tier 4+23.5points ahead
- Short factual questions, answered correctlySimpleQA Verified+17.9points ahead
- Abstract visual puzzles that people can solveARC-AGI-2+11.7points ahead
Where GPT-5.4 pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 71 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGPT-5.5
- Terminal-Bench 2.1GPT-5.5 by 678.384.3+6
- LMArena · WebDevGPT-5.5 by 48 rating points14651513+48 rating
- LiveBench · CodingGPT-5.5 by 4.777.582.2+4.7
- Terminal-Bench HardGPT-5.5 by 357.660.6+3
- SWE-bench Protie57.758.6tie
- SciCodetie56.655.8tie
- LiveBench · Agentic Codingtie53.854tie
AgentsGPT-5.5
- AA IT-Bench SREGPT-5.5 by 11.334.545.8+11.3
- MCP AtlasGPT-5.5 by 8.167.275.3+8.1
- GDPValGPT-5.5 by 5.337.442.7+5.3
- AA ApexAgentsGPT-5.5 by 4.433.337.7+4.4
- OSWorld-VerifiedGPT-5.5 by 3.77578.7+3.7
- BrowseCompGPT-5.5 by 1.782.784.4+1.7
- τ-Bench V3 · Bankingtie39.639tie
ReasoningGPT-5.5
- ARC-AGI-2GPT-5.5 by 11.773.385+11.7
- CritPtGPT-5.5 by 3.723.427.1+3.7
- Humanity's Last ExamGPT-5.5 by 2.143.745.8+2.1
- LiveBench · ReasoningGPT-5.5 by 1.688.189.7+1.6
- GPQA DiamondGPT-5.5 by 1.59293.5+1.5
- ARC-AGI-3tie0.20.4tie
FactsGPT-5.5
- SimpleQA VerifiedGPT-5.5 by 17.945.163+17.9
- AA-Omniscience · AccuracyGPT-5.5 by 7.150.958+7.1
- AA-Omniscience · Non-hallucinationGPT-5.5 by 2.78.311+2.7
Images and chartsEven
- MMMU-ProGPT-5.5 by 1.578.479.9+1.5
- LMArena · VisionGPT-5.4 by 10 rating points13021292+10 rating
MathGPT-5.5
- FrontierMath Tier 4GPT-5.5 by 23.54972.5+23.5
- FrontierMath Tiers 1-3 (v2)GPT-5.5 by 6.778.685.3+6.7
- USAMO 2026GPT-5.5 by 395.298.2+3
- LiveBench · MathematicsGPT-5.5 by 1.794.295.9+1.7
- AIME 2026tie99.2100tie
- HMMT Feb. 2026tie97.798.5tie
Following instructionsGPT-5.5
- IFBenchGPT-5.5 by 273.975.9+2
- LiveBench · Instruction Followingtie70.270.7tie
Other results37 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)GPT-5.5 by 309.739.7+30
- CyberGymGPT-5.5 by 15.566.381.8+15.5
- DeepSWE 1.1GPT-5.5 by 155267+15
- AA-OmniscienceGPT-5.5 by 14.75.820.5+14.7
- FrontierMath (overall)GPT-5.4 by 12.247.635.4+12.2
- GDPval-AA (Elo)GPT-5.5 by 95 rating points16741769+95 rating
- NL2RepoGPT-5.5 by 9.441.350.7+9.4
- frontiermath_tier_4_v1GPT-5.5 by 8.327.135.4+8.3
- Terminal-Bench 2.0GPT-5.5 by 7.675.182.7+7.6
- AA Agentic IndexGPT-5.4 by 6.944.237.3+6.9
- BabyVisionGPT-5.5 by 6.249.755.9+6.2
- livebench_languageGPT-5.5 by 4.882.687.4+4.8
- vectara_avg_summary_lengthGPT-5.5 by 47.9 rating points81.7129.6+47.9 rating
- Finance AgentGPT-5.5 by 45660+4
- Artificial Analysis Coding IndexGPT-5.5 by 3.871.174.9+3.8
- HealthBench Professional length-adjustedGPT-5.5 by 3.748.151.8+3.7
- OfficeQA ProGPT-5.5 by 351.154.1+3
- HealthBench length-adjustedGPT-5.5 by 2.55456.5+2.5
- HealthBench Hard length-adjustedGPT-5.5 by 2.429.131.5+2.4
- matharena_visual_math_overallGPT-5.5 by 2.492.594.9+2.4
- livebench_data_analysisGPT-5.5 by 2.379.381.6+2.3
- Hard-negative protein binding predictionGPT-5.4 by 23.51.5+2
- Legal Agent BenchmarkGPT-5.5 by 1.70.42.1+1.7
- ARC-AGI-1GPT-5.5 by 1.393.795+1.3
- DNA sequence design for transcription factor bindingGPT-5.5 by 112.813.8+1
- Tool-DecathlonGPT-5.5 by 154.655.6+1
- ToolathlonGPT-5.5 by 154.655.6+1
- HealthBench Consensus length-adjustedtie96.395.6tie
- HMMT Nov. 2025tie95.896.5tie
- AA Intelligencetie3938.4tie
- LiveBenchtie80.380.7tie
- HLE (with tools)tie52.152.2tie
- vectara_answer_ratetie99.9100tie
- 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-5.4 or GPT-5.5?
GPT-5.5 wins seven of the eight areas we test: coding, agents, reasoning, facts, math, long documents and following instructions. GPT-5.4 wins none, but costs 50% less. They are level on images and charts.
Which is better for coding?
GPT-5.5. It wins 4 of the 7 coding tests both models report; GPT-5.4 wins none, and 3 are ties.
Which is cheaper?
GPT-5.4 costs $2.50 per million input tokens and $15.00 per million output tokens; GPT-5.5 costs $5.00 and $30.00. That makes GPT-5.4 about 50% cheaper for the same work.
How do you compare the two?
We use the 71 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 37 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.