DeepSeek-V3.2 vs DeepSeek-V4-Pro
Wins 0 of 7 areas
—
Wins 6 of 7 areas
Coding · Agents · Reasoning · Math · Long documents · Following instructions
DeepSeek-V4-Pro is the stronger all-rounder.DeepSeek-V3.2 is cheaper.
Scores updated · 42 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.
- CodingWriting and fixing software08DeepSeek-V4-Pro8 of 8 tests
- AgentsCarrying out multi-step tasks on its own05DeepSeek-V4-Pro5 of 5 tests
- ReasoningHard problems that need careful thinking03DeepSeek-V4-Pro3 of 3 tests
- MathCompetition and research-level math03DeepSeek-V4-Pro3 of 3 tests
- Long documentsFinding answers in very long texts01DeepSeek-V4-Pro1 of 1 test
- Following instructionsDoing exactly what it is asked01DeepSeek-V4-Pro1 of 1 test
- FactsGetting facts right instead of making them up22Even2 each
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 46% 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 DeepSeek-V4-Pro pulls ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+39.8points ahead
- Finds hard-to-locate facts by browsing the webBrowseComp+32points ahead
- Real work tasks from 44 professionsGDPVal+23.1points ahead
Where DeepSeek-V3.2 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+11.4points ahead
Every test, side by side
All 42 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V4-Pro
- SWE-bench ProDeepSeek-V4-Pro by 39.815.655.4+39.8
- Terminal-Bench 2.1DeepSeek-V4-Pro by 17.246.864+17.2
- SciCodeDeepSeek-V4-Pro by 11.938.950.8+11.9
- Terminal-Bench HardDeepSeek-V4-Pro by 10.635.646.2+10.6
- LMArena · WebDevDeepSeek-V4-Pro by 102 rating points13621464+102 rating
- LiveCodeBench v6DeepSeek-V4-Pro by 9.283.392.5+9.2
- SWE-bench VerifiedDeepSeek-V4-Pro by 7.573.180.6+7.5
- SWE-bench MultilingualDeepSeek-V4-Pro by 670.276.2+6
AgentsDeepSeek-V4-Pro
- BrowseCompDeepSeek-V4-Pro by 3251.483.4+32
- GDPValDeepSeek-V4-Pro by 23.19.832.9+23.1
- MCP AtlasDeepSeek-V4-Pro by 11.462.273.6+11.4
- τ-Bench V3 · BankingDeepSeek-V4-Pro by 11.318.830.1+11.3
- AA ApexAgentsDeepSeek-V4-Pro by 9.814.524.3+9.8
ReasoningDeepSeek-V4-Pro
- Humanity's Last ExamDeepSeek-V4-Pro by 12.924.637.5+12.9
- CritPtDeepSeek-V4-Pro by 102.912.9+10
- GPQA DiamondDeepSeek-V4-Pro by 4.88488.8+4.8
FactsEven
- SimpleQA VerifiedDeepSeek-V4-Pro by 18.727.546.2+18.7
- AA-Omniscience · Non-hallucinationDeepSeek-V3.2 by 11.417.35.9+11.4
- AA-Omniscience · AccuracyDeepSeek-V4-Pro by 103343+10
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3.2 by 2.36.38.6+2.3
MathDeepSeek-V4-Pro
- IMOAnswerBenchDeepSeek-V4-Pro by 11.578.389.8+11.5
- HMMT Feb. 2026DeepSeek-V4-Pro by 11.184.195.2+11.1
- AIME 2026DeepSeek-V4-Pro by 2.594.296.7+2.5
Following instructionsDeepSeek-V4-Pro
- IFBenchDeepSeek-V4-Pro by 15.860.776.5+15.8
Other results17 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.
- CyberGymDeepSeek-V4-Pro by 6617.383.3+66
- Terminal-Bench 2.0DeepSeek-V4-Pro by 21.546.467.9+21.5
- HLE (with tools)DeepSeek-V4-Pro by 19.240.860+19.2
- Tool-DecathlonDeepSeek-V4-Pro by 17.635.252.8+17.6
- ToolathlonDeepSeek-V4-Pro by 16.635.251.8+16.6
- Artificial Analysis Coding IndexDeepSeek-V4-Pro by 15.244.259.4+15.2
- AA-OmniscienceDeepSeek-V4-Pro by 11.8-22.5-10.7+11.8
- LiveCodeBenchDeepSeek-V4-Pro by 10.283.393.5+10.2
- AA Agentic IndexDeepSeek-V4-Pro by 9.418.327.7+9.4
- vectara_avg_summary_lengthDeepSeek-V4-Pro by 91.8 rating points62153.8+91.8 rating
- AA IntelligenceDeepSeek-V4-Pro by 8.921.530.4+8.9
- GPQA (unspecified)DeepSeek-V4-Pro by 7.979.987.8+7.9
- τ²-Bench Telecom (AA run)DeepSeek-V4-Pro by 5.690.696.2+5.6
- vectara_answer_rateDeepSeek-V4-Pro by 4.692.697.2+4.6
- HMMT Nov. 2025DeepSeek-V4-Pro by 4.49094.4+4.4
- MMLU-ProDeepSeek-V4-Pro by 2.58587.5+2.5
- vectara_factual_consistencyDeepSeek-V3.2 by 2.393.791.4+2.3
Questions people ask
Which is better, DeepSeek-V3.2 or DeepSeek-V4-Pro?
DeepSeek-V4-Pro wins six of the seven areas where both have results: coding, agents, reasoning, math, long documents and following instructions. DeepSeek-V3.2 wins none, but costs 46% less. They are level on facts.
Which is better for coding?
DeepSeek-V4-Pro. It wins 8 of the 8 coding tests both models report; DeepSeek-V3.2 wins none.
Which is cheaper?
DeepSeek-V3.2 costs $0.28 per million input tokens and $0.42 per million output tokens; DeepSeek-V4-Pro costs $0.43 and $0.87. That makes DeepSeek-V3.2 about 46% cheaper for the same work.
How do you compare the two?
We use the 42 benchmark tests both models have published scores on. The verdict counts the 25 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 17 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.