DeepSeek-V3 vs DeepSeek-V3.2
Wins 0 of 6 areas
—
Wins 6 of 6 areas
Coding · Agents · Reasoning · Facts · Long documents · Following instructions
DeepSeek-V3.2 is the stronger all-rounder.
Scores updated · 30 tests both models report · How we compare
Where each one wins
Tests won in each of the six 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 software04DeepSeek-V3.24 of 5 tests · 1 tie
- ReasoningHard problems that need careful thinking03DeepSeek-V3.23 of 3 tests
- FactsGetting facts right instead of making them up02DeepSeek-V3.22 of 3 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own02DeepSeek-V3.22 of 2 tests
- Long documentsFinding answers in very long texts02DeepSeek-V3.22 of 2 tests
- Following instructionsDoing exactly what it is asked01DeepSeek-V3.21 of 1 test
Following instructions 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-V3.2 costs 39% 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-V3.2 pulls ahead
- Recent programming contest problemsLiveCodeBench v6+36.4points ahead
- Reasons across sets of long documentsAA-LCR+32.6points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+31.1points ahead
Where DeepSeek-V3 pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 30 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3.2
- LiveCodeBench v6DeepSeek-V3.2 by 36.446.983.3+36.4
- SWE-bench VerifiedDeepSeek-V3.2 by 31.14273.1+31.1
- Terminal-Bench 2.1DeepSeek-V3.2 by 29.916.946.8+29.9
- Terminal-Bench HardDeepSeek-V3.2 by 20.415.235.6+20.4
- SciCodetie3938.9tie
AgentsDeepSeek-V3.2
- τ-Bench V3 · BankingDeepSeek-V3.2 by 14.14.718.8+14.1
- GDPValDeepSeek-V3.2 by 9.809.8+9.8
ReasoningDeepSeek-V3.2
- Humanity's Last ExamDeepSeek-V3.2 by 19.94.724.6+19.9
- GPQA DiamondDeepSeek-V3.2 by 18.565.584+18.5
- CritPtDeepSeek-V3.2 by 2.902.9+2.9
FactsDeepSeek-V3.2
- AA-Omniscience · AccuracyDeepSeek-V3.2 by 7.525.433+7.5
- AA-Omniscience · Non-hallucinationDeepSeek-V3.2 by 3.214.117.3+3.2
- Vectara HHEM hallucination ratelower is bettertie6.16.3tie
Long documentsDeepSeek-V3.2
- AA-LCRDeepSeek-V3.2 by 32.640.773.3+32.6
- LongBench v2DeepSeek-V3.2 by 11.148.759.8+11.1
Following instructionsDeepSeek-V3.2
- IFBenchDeepSeek-V3.2 by 19.74160.7+19.7
Other results14 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.
- HMMT Feb. 2025DeepSeek-V3.2 by 63.329.292.5+63.3
- HMMT 2025DeepSeek-V3.2 by 62.727.590.2+62.7
- τ²-BenchDeepSeek-V3.2 by 47.832.580.3+47.8
- τ²-Bench Telecom (AA run)DeepSeek-V3.2 by 43.547.190.6+43.5
- AIME 2025DeepSeek-V3.2 by 41.851.393.1+41.8
- LiveCodeBenchDeepSeek-V3.2 by 34.149.283.3+34.1
- Artificial Analysis Coding IndexDeepSeek-V3.2 by 21.22344.2+21.2
- vectara_avg_summary_lengthDeepSeek-V3 by 19.781.762+19.7
- AA-OmniscienceDeepSeek-V3.2 by 18.2-40.7-22.5+18.2
- AA IntelligenceDeepSeek-V3.2 by 11.89.721.5+11.8
- MMLU-ProDeepSeek-V3.2 by 9.175.985+9.1
- vectara_answer_rateDeepSeek-V3 by 4.997.592.6+4.9
- MMLU-ReduxDeepSeek-V3.2 by 4.689.193.7+4.6
- vectara_factual_consistencytie93.993.7tie
Questions people ask
Which is better, DeepSeek-V3 or DeepSeek-V3.2?
DeepSeek-V3.2 wins all six areas where both have results: coding, agents, reasoning, facts, long documents and following instructions. DeepSeek-V3 wins none.
Which is better for coding?
DeepSeek-V3.2. It wins 4 of the 5 coding tests both models report; DeepSeek-V3 wins none, and 1 is a tie.
Which is cheaper?
DeepSeek-V3 costs $0.24 per million input tokens and $0.90 per million output tokens; DeepSeek-V3.2 costs $0.28 and $0.42. That makes DeepSeek-V3.2 about 39% cheaper for the same work.
How do you compare the two?
We use the 30 benchmark tests both models have published scores on. The verdict counts the 16 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 14 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.