DeepSeek-V3.2 vs Gemini 3 Pro
Wins 0 of 7 areas
—
Wins 6 of 7 areas
Coding · Agents · Reasoning · Math · Long documents · Following instructions
Gemini 3 Pro is the stronger all-rounder.DeepSeek-V3.2 is cheaper.
Scores updated · 62 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 software16Gemini 3 Pro6 of 7 tests
- ReasoningHard problems that need careful thinking04Gemini 3 Pro4 of 4 tests
- AgentsCarrying out multi-step tasks on its own13Gemini 3 Pro3 of 4 tests
- Long documentsFinding answers in very long texts02Gemini 3 Pro2 of 2 tests
- MathCompetition and research-level math12Gemini 3 Pro2 of 3 tests
- Following instructionsDoing exactly what it is asked01Gemini 3 Pro1 of 1 test
- FactsGetting facts right instead of making them up22Even2 each
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 95% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Gemini 3 Pro pulls ahead
- Short factual questions, answered correctlySimpleQA Verified+45.4points ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+27.7points ahead
- Abstract visual puzzles that people can solveARC-AGI-2+27.1points ahead
Where DeepSeek-V3.2 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+8.8points ahead
- Multi-step tasks using many real software toolsMCP Atlas+8.1points ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+5.2points ahead
Every test, side by side
All 62 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGemini 3 Pro
- SWE-bench ProGemini 3 Pro by 27.715.643.3+27.7
- SciCodeGemini 3 Pro by 17.238.956.1+17.2
- LMArena · WebDevGemini 3 Pro by 78 rating points13621440+78 rating
- Terminal-Bench HardGemini 3 Pro by 6.135.641.7+6.1
- SWE-bench MultilingualDeepSeek-V3.2 by 5.270.265+5.2
- LiveCodeBench v6Gemini 3 Pro by 4.183.387.4+4.1
- SWE-bench VerifiedGemini 3 Pro by 3.173.176.2+3.1
AgentsGemini 3 Pro
- GDPValGemini 3 Pro by 24.49.834.2+24.4
- MCP AtlasDeepSeek-V3.2 by 8.162.254.1+8.1
- BrowseCompGemini 3 Pro by 7.851.459.2+7.8
- AA ApexAgentsGemini 3 Pro by 3.914.518.4+3.9
ReasoningGemini 3 Pro
- ARC-AGI-2Gemini 3 Pro by 27.1431.1+27.1
- Humanity's Last ExamGemini 3 Pro by 15.124.639.7+15.1
- GPQA DiamondGemini 3 Pro by 6.88490.8+6.8
- CritPtGemini 3 Pro by 6.22.99.1+6.2
FactsEven
- SimpleQA VerifiedGemini 3 Pro by 45.427.572.9+45.4
- AA-Omniscience · AccuracyGemini 3 Pro by 22.83355.8+22.8
- AA-Omniscience · Non-hallucinationDeepSeek-V3.2 by 8.817.38.5+8.8
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3.2 by 7.36.313.6+7.3
MathGemini 3 Pro
- IMOAnswerBenchGemini 3 Pro by 4.878.383.1+4.8
- AIME 2026DeepSeek-V3.2 by 2.594.291.7+2.5
- HMMT Feb. 2026Gemini 3 Pro by 2.384.186.4+2.3
Long documentsGemini 3 Pro
- LongBench v2Gemini 3 Pro by 8.459.868.2+8.4
- AA-LCRGemini 3 Pro by 2.773.376+2.7
Following instructionsGemini 3 Pro
- IFBenchGemini 3 Pro by 9.760.770.4+9.7
Other results37 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.
- Vending-Bench 2Gemini 3 Pro by 546782 rating points1034547816+546782 rating
- AA-OmniscienceGemini 3 Pro by 37.8-22.515.3+37.8
- MRCRGemini 3 Pro by 34.255.589.7+34.2
- AA Agentic IndexGemini 3 Pro by 33.718.352+33.7
- OJ-Bench (cpp)Gemini 3 Pro by 30.338.268.5+30.3
- WideSearch (item-f1)Gemini 3 Pro by 24.532.557+24.5
- Arena-Hard (Hard Prompt)Gemini 3 Pro by 19.253.472.6+19.2
- ARC-AGI-1Gemini 3 Pro by 185775+18
- SWT-benchGemini 3 Pro by 17.76279.7+17.7
- τ²-BenchGemini 3 Pro by 17.780.398+17.7
- frontiermath_tier_4_v1Gemini 3 Pro by 16.72.118.8+16.7
- Multi-SWE-BenchGemini 3 Pro by 12.130.642.7+12.1
- FinSearchCompT2&T3DeepSeek-V3.2 by 9.259.149.9+9.2
- LiveCodeBenchGemini 3 Pro by 8.783.392+8.7
- BrowseComp (context management)DeepSeek-V3.2 by 8.467.659.2+8.4
- browsecomp_with_context_managerDeepSeek-V3.2 by 8.467.659.2+8.4
- Terminal-Bench 2.0Gemini 3 Pro by 7.846.454.2+7.8
- vectara_factual_consistencyDeepSeek-V3.2 by 7.393.786.4+7.3
- AIME 2025Gemini 3 Pro by 6.993.1100+6.9
- vectara_answer_rateGemini 3 Pro by 6.892.699.4+6.8
- AIME25 no toolsGemini 3 Pro by 6.789.396+6.7
- AA IntelligenceGemini 3 Pro by 6.521.528+6.5
- SWE-PerfGemini 3 Pro by 5.60.96.5+5.6
- t2-benchGemini 3 Pro by 5.280.285.4+5.2
- MMLU-ProGemini 3 Pro by 5.18590.1+5.1
- HLE (with tools)Gemini 3 Pro by 540.845.8+5
- HMMT Feb. 2025Gemini 3 Pro by 592.597.5+5
- Arena-Hard (Creative Writing)Gemini 3 Pro by 4.888.893.6+4.8
- Seal-0DeepSeek-V3.2 by 449.545.5+4
- vectara_avg_summary_lengthGemini 3 Pro by 39.9 rating points62101.9+39.9 rating
- τ²-Bench Telecom (AA run)DeepSeek-V3.2 by 3.590.687.1+3.5
- HMMT Nov. 2025Gemini 3 Pro by 3.39093.3+3.3
- OctoCodingbenchDeepSeek-V3.2 by 3.12622.9+3.1
- Artificial Analysis Coding IndexGemini 3 Pro by 2.344.246.5+2.3
- DeepSearchQA (F1)Gemini 3 Pro by 2.360.963.2+2.3
- AIME 2026 I (Tools-allowed)DeepSeek-V3.2 by 2.192.790.6+2.1
- ToolathlonGemini 3 Pro by 1.235.236.4+1.2
Questions people ask
Which is better, DeepSeek-V3.2 or Gemini 3 Pro?
Gemini 3 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 95% less. They are level on facts.
Which is better for coding?
Gemini 3 Pro. It wins 6 of the 7 coding tests both models report; DeepSeek-V3.2 wins 1.
Which is cheaper?
DeepSeek-V3.2 costs $0.28 per million input tokens and $0.42 per million output tokens; Gemini 3 Pro costs $2.00 and $12.00. That makes DeepSeek-V3.2 about 95% cheaper for the same work.
How do you compare the two?
We use the 62 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 37 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.