DeepSeek-V4-Pro vs GPT-5.4
Wins 0 of 7 areas
—
Wins 6 of 7 areas
Coding · Agents · Reasoning · Facts · Math · Long documents
GPT-5.4 is the stronger all-rounder.DeepSeek-V4-Pro is cheaper.
Scores updated · 64 tests both models report · How we compare
Where each one wins
Tests won in each of the seven areas where both have results. 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 math07GPT-5.47 of 7 tests
- CodingWriting and fixing software16GPT-5.46 of 8 tests · 1 tie
- ReasoningHard problems that need careful thinking04GPT-5.44 of 4 tests
- FactsGetting facts right instead of making them up13GPT-5.43 of 4 tests
- AgentsCarrying out multi-step tasks on its own23GPT-5.43 of 6 tests · 1 tie
- Long documentsFinding answers in very long texts01GPT-5.41 of 1 test
- 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.
DeepSeek-V4-Pro costs 93% 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
- Extremely hard research-level math problemsFrontierMath Tier 4+46.6points ahead
- Proof problems from the US Math OlympiadUSAMO 2026+34.5points ahead
- Unpublished advanced math problemsFrontierMath Tiers 1-3 (v2)+33.3points ahead
Where DeepSeek-V4-Pro pulls ahead
- Multi-step tasks using many real software toolsMCP Atlas+6.4points ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+4.5points ahead
- Finds the root cause of IT system incidentsAA IT-Bench SRE+3.8points ahead
Every test, side by side
All 64 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 14.36478.3+14.3
- Terminal-Bench HardGPT-5.4 by 11.446.257.6+11.4
- LiveBench · Agentic CodingGPT-5.4 by 11.242.653.8+11.2
- LiveBench · CodingGPT-5.4 by 7.57077.5+7.5
- SciCodeGPT-5.4 by 5.850.856.6+5.8
- SWE-bench MultilingualDeepSeek-V4-Pro by 4.576.271.7+4.5
- SWE-bench ProGPT-5.4 by 2.355.457.7+2.3
- LMArena · WebDevtie14641465tie
AgentsGPT-5.4
- τ-Bench V3 · BankingGPT-5.4 by 9.530.139.6+9.5
- AA ApexAgentsGPT-5.4 by 924.333.3+9
- MCP AtlasDeepSeek-V4-Pro by 6.473.667.2+6.4
- GDPValGPT-5.4 by 4.532.937.4+4.5
- AA IT-Bench SREDeepSeek-V4-Pro by 3.838.334.5+3.8
- BrowseComptie83.482.7tie
ReasoningGPT-5.4
- CritPtGPT-5.4 by 10.512.923.4+10.5
- Humanity's Last ExamGPT-5.4 by 6.237.543.7+6.2
- LiveBench · ReasoningGPT-5.4 by 5.482.788.1+5.4
- GPQA DiamondGPT-5.4 by 3.288.892+3.2
FactsGPT-5.4
- AA-Omniscience · AccuracyGPT-5.4 by 7.94350.9+7.9
- AA-Omniscience · Non-hallucinationGPT-5.4 by 2.45.98.3+2.4
- Vectara HHEM hallucination ratelower is betterGPT-5.4 by 1.68.67+1.6
- SimpleQA VerifiedDeepSeek-V4-Pro by 1.146.245.1+1.1
MathGPT-5.4
- FrontierMath Tier 4GPT-5.4 by 46.62.449+46.6
- USAMO 2026GPT-5.4 by 34.560.795.2+34.5
- FrontierMath Tiers 1-3 (v2)GPT-5.4 by 33.345.378.6+33.3
- LiveBench · MathematicsGPT-5.4 by 3.590.794.2+3.5
- AIME 2026GPT-5.4 by 2.596.799.2+2.5
- HMMT Feb. 2026GPT-5.4 by 2.595.297.7+2.5
- IMOAnswerBenchGPT-5.4 by 1.689.891.4+1.6
Following instructionsEven
- LiveBench · Instruction FollowingGPT-5.4 by 7.862.470.2+7.8
- IFBenchDeepSeek-V4-Pro by 2.676.573.9+2.6
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.
- Apex (Pass@1)GPT-5.4 by 26.727.454.1+26.7
- Codeforces (Rating)DeepSeek-V4-Pro by 180 rating points33483168+180 rating
- CyberGymDeepSeek-V4-Pro by 1783.366.3+17
- AA Agentic IndexGPT-5.4 by 16.527.744.2+16.5
- AA-OmniscienceGPT-5.4 by 16.5-10.75.8+16.5
- GDPval-AA (Elo)GPT-5.4 by 120 rating points15541674+120 rating
- Artificial Analysis Coding IndexGPT-5.4 by 11.759.471.1+11.7
- AgentWorldBench - SearchGPT-5.4 by 9.727.637.3+9.7
- τ²-Bench Telecom (AA run)DeepSeek-V4-Pro by 9.196.287.1+9.1
- AA IntelligenceGPT-5.4 by 8.630.439+8.6
- HLE (with tools)DeepSeek-V4-Pro by 86052+8
- Apex Shortlist (Pass@1)DeepSeek-V4-Pro by 7.485.578.1+7.4
- vectara_avg_summary_lengthDeepSeek-V4-Pro by 72.1 rating points153.881.7+72.1 rating
- Terminal-Bench 2.0GPT-5.4 by 7.267.975.1+7.2
- AgentWorldBench - SWEGPT-5.4 by 6.959.466.3+6.9
- AgentWorldBench - MCPGPT-5.4 by 6.863.370.1+6.8
- LiveBenchGPT-5.4 by 6.773.680.3+6.7
- MCPAtlas Public (Pass@1)DeepSeek-V4-Pro by 6.473.667.2+6.4
- AgentWorldBench - OverallGPT-5.4 by 5.35358.3+5.3
- AgentWorldBench - OSGPT-5.4 by 4.963.768.6+4.9
- AgentWorldBench - AndroidGPT-5.4 by 4.855.260+4.8
- livebench_data_analysisGPT-5.4 by 4.874.579.3+4.8
- livebench_languageGPT-5.4 by 4.578.182.6+4.5
- NL2RepoGPT-5.4 by 2.838.541.3+2.8
- ToolathlonGPT-5.4 by 2.851.854.6+2.8
- vectara_answer_rateGPT-5.4 by 2.797.299.9+2.7
- AgentWorldBench - TerminalGPT-5.4 by 2.451.353.7+2.4
- Tool-DecathlonGPT-5.4 by 1.852.854.6+1.8
- vectara_factual_consistencyGPT-5.4 by 1.691.493+1.6
- AgentWorldBench - WebGPT-5.4 by 1.550.351.8+1.5
- HMMT Nov. 2025GPT-5.4 by 1.494.495.8+1.4
- MMLU-Protie87.587.5tie
Questions people ask
Which is better, DeepSeek-V4-Pro or GPT-5.4?
GPT-5.4 wins six of the seven areas where both have results: coding, agents, reasoning, facts, math and long documents. DeepSeek-V4-Pro wins none, but costs 93% less. They are level on following instructions.
Which is better for coding?
GPT-5.4. It wins 6 of the 8 coding tests both models report; DeepSeek-V4-Pro wins 1, and 1 is a tie.
Which is cheaper?
DeepSeek-V4-Pro costs $0.43 per million input tokens and $0.87 per million output tokens; GPT-5.4 costs $2.50 and $15.00. That makes DeepSeek-V4-Pro about 93% cheaper for the same work.
How do you compare the two?
We use the 64 benchmark tests both models have published scores on. The verdict counts the 32 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.