GPT-5.4 vs Qwen3.5 397B A17B
Wins 7 of 8 areas
Coding · Agents · Reasoning · Facts · Images and charts · Math · Long documents
Wins 1 of 8 areas
Following instructions
GPT-5.4 is the stronger all-rounder.Qwen3.5 397B A17B is cheaper and better at following instructions.
Scores updated · 46 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 software60GPT-5.46 of 6 tests
- MathCompetition and research-level math50GPT-5.45 of 5 tests
- AgentsCarrying out multi-step tasks on its own40GPT-5.44 of 5 tests · 1 tie
- ReasoningHard problems that need careful thinking30GPT-5.43 of 3 tests
- Images and chartsUnderstanding pictures, charts and video20GPT-5.42 of 2 tests
- FactsGetting facts right instead of making them up21GPT-5.42 of 3 tests
- Long documentsFinding answers in very long texts10GPT-5.41 of 1 test
- Following instructionsDoing exactly what it is asked01Qwen3.5 397B A17B1 of 1 test
Long documents and following instructions rest on a single test 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.
Qwen3.5 397B A17B costs 76% 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-5.4 pulls ahead
- Proof problems from the US Math OlympiadUSAMO 2026+58.9points ahead
- Unpublished advanced math problemsFrontierMath Tiers 1-3 (v2)+47.4points ahead
- Command-line tasks in a real terminalTerminal-Bench 2.1+27points ahead
Where Qwen3.5 397B A17B pulls ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+4.9points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+2.8points ahead
Every test, side by side
All 46 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGPT-5.4
- Terminal-Bench 2.1GPT-5.4 by 2778.351.3+27
- Terminal-Bench HardGPT-5.4 by 16.757.640.9+16.7
- SciCodeGPT-5.4 by 11.856.644.8+11.8
- SWE-bench ProGPT-5.4 by 6.857.750.9+6.8
- LMArena · WebDevGPT-5.4 by 65 rating points14651400+65 rating
- SWE-bench MultilingualGPT-5.4 by 2.471.769.3+2.4
AgentsGPT-5.4
- τ-Bench V3 · BankingGPT-5.4 by 26.239.613.4+26.2
- GDPValGPT-5.4 by 22.637.414.8+22.6
- AA ApexAgentsGPT-5.4 by 1833.315.3+18
- BrowseCompGPT-5.4 by 13.782.769+13.7
- AA IT-Bench SREtie34.534.1tie
ReasoningGPT-5.4
- CritPtGPT-5.4 by 21.723.41.7+21.7
- Humanity's Last ExamGPT-5.4 by 14.743.729+14.7
- GPQA DiamondGPT-5.4 by 2.79289.3+2.7
FactsGPT-5.4
- AA-Omniscience · AccuracyGPT-5.4 by 20.150.930.8+20.1
- SimpleQA VerifiedGPT-5.4 by 19.145.126+19.1
- AA-Omniscience · Non-hallucinationQwen3.5 397B A17B by 2.88.311.1+2.8
Images and chartsGPT-5.4
- LMArena · VisionGPT-5.4 by 38 rating points13021264+38 rating
- MMMU-ProGPT-5.4 by 1.178.477.3+1.1
MathGPT-5.4
- USAMO 2026GPT-5.4 by 58.995.236.3+58.9
- FrontierMath Tiers 1-3 (v2)GPT-5.4 by 47.478.631.2+47.4
- IMOAnswerBenchGPT-5.4 by 10.591.480.9+10.5
- HMMT Feb. 2026GPT-5.4 by 9.897.787.9+9.8
- AIME 2026GPT-5.4 by 7.999.291.3+7.9
Following instructionsQwen3.5 397B A17B
- IFBenchQwen3.5 397B A17B by 4.973.978.8+4.9
Other results20 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.
- AA-OmniscienceGPT-5.4 by 36.55.8-30.8+36.5
- AA Agentic IndexGPT-5.4 by 33.644.210.6+33.6
- Artificial Analysis Coding IndexGPT-5.4 by 22.971.148.2+22.9
- Terminal-Bench 2.0GPT-5.4 by 22.675.152.5+22.6
- AA IntelligenceGPT-5.4 by 20.63918.4+20.6
- MCPMarkGPT-5.4 by 16.462.546.1+16.4
- ToolathlonGPT-5.4 by 16.354.638.3+16.3
- NL2RepoGPT-5.4 by 9.141.332.2+9.1
- τ²-Bench Telecom (AA run)Qwen3.5 397B A17B by 8.587.195.6+8.5
- AgentWorldBench - OSGPT-5.4 by 7.768.660.9+7.7
- AgentWorldBench - SearchGPT-5.4 by 6.537.330.8+6.5
- AgentWorldBench - AndroidGPT-5.4 by 5.16054.9+5.1
- HLE (with tools)GPT-5.4 by 3.852.148.3+3.8
- AgentWorldBench - OverallGPT-5.4 by 3.658.354.7+3.6
- AgentWorldBench - WebGPT-5.4 by 3.251.848.5+3.2
- HMMT Nov. 2025GPT-5.4 by 3.195.892.7+3.1
- AgentWorldBench - SWEGPT-5.4 by 1.966.364.4+1.9
- AgentWorldBench - MCPGPT-5.4 by 1.870.168.3+1.8
- AgentWorldBench - TerminalQwen3.5 397B A17B by 1.653.755.3+1.6
- MMLU-Protie87.587.8tie
Questions people ask
Which is better, GPT-5.4 or Qwen3.5 397B A17B?
GPT-5.4 wins seven of the eight areas we test: coding, agents, reasoning, facts, images and charts, math and long documents. Qwen3.5 397B A17B wins following instructions, and costs 76% less.
Which is better for coding?
GPT-5.4. It wins 6 of the 6 coding tests both models report; Qwen3.5 397B A17B wins none.
Which is cheaper?
GPT-5.4 costs $2.50 per million input tokens and $15.00 per million output tokens; Qwen3.5 397B A17B costs $0.60 and $3.60. That makes Qwen3.5 397B A17B about 76% cheaper for the same work.
How do you compare the two?
We use the 46 benchmark tests both models have published scores on. The verdict counts the 26 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 20 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.