DeepSeek-V4-Pro vs GPT-5.2
Wins 3 of 7 areas
Agents · Reasoning · Following instructions
Wins 2 of 7 areas
Math · Long documents
DeepSeek-V4-Pro wins more areas, narrowly.GPT-5.2 is better at math.
Scores updated · 50 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.
- AgentsCarrying out multi-step tasks on its own41DeepSeek-V4-Pro4 of 5 tests
- ReasoningHard problems that need careful thinking21DeepSeek-V4-Pro2 of 5 tests · 2 ties
- Following instructionsDoing exactly what it is asked10DeepSeek-V4-Pro1 of 2 tests · 1 tie
- MathCompetition and research-level math15GPT-5.25 of 6 tests
- Long documentsFinding answers in very long texts01GPT-5.21 of 1 test
- CodingWriting and fixing software33Even3 each · 2 ties
- 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.
DeepSeek-V4-Pro costs 92% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where DeepSeek-V4-Pro pulls ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+25.5points ahead
- Customer-service tasks in a simulated bankτ-Bench V3 · Banking+19points ahead
- Finds hard-to-locate facts by browsing the webBrowseComp+17.6points ahead
Where GPT-5.2 pulls ahead
- Extremely hard research-level math problemsFrontierMath Tier 4+29.3points ahead
- Unpublished advanced math problemsFrontierMath Tiers 1-3 (v2)+22.1points ahead
- Real work tasks from 44 professionsGDPVal+15.4points ahead
Every test, side by side
All 50 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingEven
- SWE-bench ProDeepSeek-V4-Pro by 25.555.429.9+25.5
- SWE-bench MultilingualDeepSeek-V4-Pro by 9.576.266.7+9.5
- LiveBench · Agentic CodingGPT-5.2 by 7.742.650.3+7.7
- LiveBench · CodingGPT-5.2 by 6.17076.1+6.1
- LMArena · WebDevDeepSeek-V4-Pro by 48 rating points14631415+48 rating
- SciCodeGPT-5.2 by 1.350.852.1+1.3
- Terminal-Bench Hardtie46.247tie
- SWE-bench Verifiedtie80.680tie
AgentsDeepSeek-V4-Pro
- τ-Bench V3 · BankingDeepSeek-V4-Pro by 1930.111.1+19
- BrowseCompDeepSeek-V4-Pro by 17.683.465.8+17.6
- GDPValGPT-5.2 by 15.432.948.3+15.4
- MCP AtlasDeepSeek-V4-Pro by 1373.660.6+13
- AA ApexAgentsDeepSeek-V4-Pro by 1.324.323+1.3
ReasoningDeepSeek-V4-Pro
- SimpleBenchDeepSeek-V4-Pro by 5.150.945.8+5.1
- GPQA DiamondGPT-5.2 by 1.588.890.3+1.5
- CritPtDeepSeek-V4-Pro by 1.312.911.6+1.3
- LiveBench · Reasoningtie82.783.2tie
- Humanity's Last Examtie37.537.7tie
FactsEven
- AA-Omniscience · Non-hallucinationGPT-5.2 by 12.95.918.8+12.9
- SimpleQA VerifiedDeepSeek-V4-Pro by 9.146.237.1+9.1
- Vectara HHEM hallucination ratelower is betterDeepSeek-V4-Pro by 2.28.610.8+2.2
- AA-Omniscience · AccuracyGPT-5.2 by 1.34344.3+1.3
MathGPT-5.2
- FrontierMath Tier 4GPT-5.2 by 29.32.431.7+29.3
- FrontierMath Tiers 1-3 (v2)GPT-5.2 by 22.145.367.4+22.1
- IMOAnswerBenchDeepSeek-V4-Pro by 3.589.886.3+3.5
- LiveBench · MathematicsGPT-5.2 by 2.590.793.2+2.5
- HMMT Feb. 2026GPT-5.2 by 1.895.297+1.8
- AIME 2026GPT-5.2 by 1.696.798.3+1.6
Following instructionsDeepSeek-V4-Pro
- IFBenchDeepSeek-V4-Pro by 1.176.575.4+1.1
- LiveBench · Instruction Followingtie62.461.8tie
Other results19 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 Agentic IndexGPT-5.2 by 32.527.760.2+32.5
- HLE (with tools)DeepSeek-V4-Pro by 14.56045.5+14.5
- Terminal-Bench 2.0DeepSeek-V4-Pro by 13.967.954+13.9
- τ²-Bench Telecom (AA run)DeepSeek-V4-Pro by 11.496.284.8+11.4
- Artificial Analysis Coding IndexDeepSeek-V4-Pro by 10.759.448.7+10.7
- AA-OmniscienceGPT-5.2 by 9.8-10.7-0.9+9.8
- GDPval-AA (Elo)DeepSeek-V4-Pro by 92 rating points15541462+92 rating
- ToolathlonDeepSeek-V4-Pro by 5.551.846.3+5.5
- GPQA (unspecified)GPT-5.2 by 4.687.892.4+4.6
- LiveCodeBenchDeepSeek-V4-Pro by 4.593.589+4.5
- livebench_data_analysisGPT-5.2 by 3.774.578.2+3.7
- vectara_avg_summary_lengthGPT-5.2 by 32.5 rating points153.8186.3+32.5 rating
- vectara_answer_rateGPT-5.2 by 2.897.2100+2.8
- vectara_factual_consistencyDeepSeek-V4-Pro by 2.291.489.2+2.2
- livebench_languageGPT-5.2 by 1.778.179.8+1.7
- HMMT Nov. 2025GPT-5.2 by 1.494.495.8+1.4
- LiveBenchGPT-5.2 by 1.273.674.8+1.2
- MMLU-Protie87.586.7tie
- AA Intelligencetie30.430.4tie
Questions people ask
Which is better, DeepSeek-V4-Pro or GPT-5.2?
DeepSeek-V4-Pro wins three of the seven areas where both have results: agents, reasoning and following instructions. GPT-5.2 wins math and long documents. They are level on coding and facts.
Which is better for coding?
Neither. They win 3 coding tests each of the 8 both models report, and 2 are ties.
Which is cheaper?
DeepSeek-V4-Pro costs $0.43 per million input tokens and $0.87 per million output tokens; GPT-5.2 costs $1.75 and $14.00. That makes DeepSeek-V4-Pro about 92% cheaper for the same work.
How do you compare the two?
We use the 50 benchmark tests both models have published scores on. The verdict counts the 31 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 19 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.