DeepSeek-V3.2 vs GPT-5
Wins 1 of 7 areas
Math
Wins 6 of 7 areas
Coding · Agents · Reasoning · Facts · Long documents · Following instructions
GPT-5 is the stronger all-rounder.DeepSeek-V3.2 is cheaper and better at math.
Scores updated · 67 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.
- ReasoningHard problems that need careful thinking04GPT-54 of 4 tests
- AgentsCarrying out multi-step tasks on its own03GPT-53 of 3 tests
- CodingWriting and fixing software35GPT-55 of 8 tests
- FactsGetting facts right instead of making them up12GPT-52 of 4 tests · 1 tie
- Long documentsFinding answers in very long texts01GPT-51 of 1 test
- Following instructionsDoing exactly what it is asked01GPT-51 of 1 test
- MathCompetition and research-level math10DeepSeek-V3.21 of 1 test
Math, 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.
DeepSeek-V3.2 costs 94% 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 pulls ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+26.2points ahead
- Short factual questions, answered correctlySimpleQA Verified+22.6points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+12.4points ahead
Where DeepSeek-V3.2 pulls ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+14.9points ahead
- Command-line tasks in a real terminalTerminal-Bench 2.1+11.6points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+3points ahead
Every test, side by side
All 67 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGPT-5
- SWE-bench ProGPT-5 by 26.215.641.8+26.2
- SWE-bench MultilingualDeepSeek-V3.2 by 14.970.255.3+14.9
- Terminal-Bench 2.1DeepSeek-V3.2 by 11.646.835.2+11.6
- LMArena · WebDevGPT-5 by 56 rating points13621418+56 rating
- SciCodeGPT-5 by 438.942.9+4
- LiveCodeBench v6GPT-5 by 3.783.387+3.7
- Terminal-Bench HardDeepSeek-V3.2 by 335.632.6+3
- SWE-bench VerifiedGPT-5 by 1.873.174.9+1.8
AgentsGPT-5
- GDPValGPT-5 by 11.29.821+11.2
- BrowseCompGPT-5 by 3.551.454.9+3.5
- τ-Bench V3 · BankingGPT-5 by 3.318.822.1+3.3
ReasoningGPT-5
- ARC-AGI-2GPT-5 by 5.949.9+5.9
- Humanity's Last ExamGPT-5 by 3.924.628.5+3.9
- CritPtGPT-5 by 2.82.95.7+2.8
- GPQA DiamondGPT-5 by 1.48485.4+1.4
FactsGPT-5
- SimpleQA VerifiedGPT-5 by 22.627.550.1+22.6
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3.2 by 8.86.315.1+8.8
- AA-Omniscience · AccuracyGPT-5 by 7.33340.3+7.3
- AA-Omniscience · Non-hallucinationtie17.317.8tie
Other results45 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.
- FinSearchComp-globalGPT-5 by 37.726.263.9+37.7
- xbench-DeepSearchGPT-5 by 22.155.777.8+22.1
- FinSearchComp-T3 (Tools-allowed)GPT-5 by 21.52748.5+21.5
- FinSearchComp-T3 w/ toolsGPT-5 by 21.52748.5+21.5
- HealthBenchGPT-5 by 20.346.967.2+20.3
- HealthBench no toolsGPT-5 by 20.346.967.2+20.3
- Arena-Hard (Hard Prompt)GPT-5 by 18.553.471.9+18.5
- OJ-Bench (cpp)GPT-5 by 1838.256.2+18
- OJ-Bench (cpp) no toolsGPT-5 by 1838.256.2+18
- ArtifactsBenchGPT-5 by 17.255.873+17.2
- BrowseComp-ZH w/ toolsGPT-5 by 15.147.963+15.1
- BrowseComp w/ toolsGPT-5 by 14.840.154.9+14.8
- AA-OmniscienceGPT-5 by 13.8-22.5-8.7+13.8
- GAIA (text only)GPT-5 by 12.963.576.4+12.9
- LiveCodeBenchV6 no toolsGPT-5 by 12.974.187+12.9
- Seal-0 w/ toolsGPT-5 by 12.938.551.4+12.9
- Terminal-Bench 2.0DeepSeek-V3.2 by 11.246.435.2+11.2
- frontiermath_tier_4_v1GPT-5 by 10.42.112.5+10.4
- vectara_avg_summary_lengthGPT-5 by 100.7 rating points62162.7+100.7 rating
- HMMT25 no toolsGPT-5 by 9.783.693.3+9.7
- vectara_factual_consistencyDeepSeek-V3.2 by 8.893.784.9+8.8
- ARC-AGI-1GPT-5 by 8.75765.7+8.7
- AA Agentic IndexGPT-5 by 8.218.326.5+8.2
- vectara_answer_rateGPT-5 by 7.392.699.9+7.3
- Artificial Analysis Coding IndexDeepSeek-V3.2 by 6.444.237.8+6.4
- Terminal-BenchGPT-5 by 6.137.743.8+6.1
- Terminal-Bench w/ simulated tools (JSON)GPT-5 by 6.137.743.8+6.1
- Frames w/ toolsGPT-5 by 5.880.286+5.8
- GPQA (unspecified)GPT-5 by 5.879.985.7+5.8
- τ²-Bench Telecom (AA run)DeepSeek-V3.2 by 5.890.684.8+5.8
- HLE (with tools)DeepSeek-V3.2 by 5.640.835.2+5.6
- AIME25 no toolsGPT-5 by 5.389.394.6+5.3
- swe_bench_bashGPT-5 by 56065+5
- HMMT Feb. 2025DeepSeek-V3.2 by 4.292.588.3+4.2
- Arena-Hard (Creative Writing)GPT-5 by 3.488.892.2+3.4
- HMMT 2025GPT-5 by 3.190.293.3+3.1
- MMLU-ProGPT-5 by 2.18587.1+2.1
- τ²-BenchGPT-5 by 2.180.382.4+2.1
- BrowseComp-ZHDeepSeek-V3.2 by 26563+2
- MMLU-ReduxGPT-5 by 1.693.795.3+1.6
- AA IntelligenceGPT-5 by 1.521.523+1.5
- AIME 2025GPT-5 by 1.593.194.6+1.5
- Longform Writing eval (Kimi K2 Thinking system card)DeepSeek-V3.2 by 1.172.571.4+1.1
- HMMT Nov. 2025tie9089.2tie
- LiveCodeBenchtie83.383.5tie
Questions people ask
Which is better, DeepSeek-V3.2 or GPT-5?
GPT-5 wins six of the seven areas where both have results: coding, agents, reasoning, facts, long documents and following instructions. DeepSeek-V3.2 wins math, and costs 94% less.
Which is better for coding?
GPT-5. It wins 5 of the 8 coding tests both models report; DeepSeek-V3.2 wins 3.
Which is cheaper?
DeepSeek-V3.2 costs $0.28 per million input tokens and $0.42 per million output tokens; GPT-5 costs $1.25 and $10.00. That makes DeepSeek-V3.2 about 94% cheaper for the same work.
How do you compare the two?
We use the 67 benchmark tests both models have published scores on. The verdict counts the 22 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 45 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.