The two are evenly matched.DeepSeek-V3 is better at coding and facts; Qwen3 235B A22B at reasoning and agents.
Scores updated · 55 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 software30DeepSeek-V33 of 4 tests · 1 tie
- FactsGetting facts right instead of making them up21DeepSeek-V32 of 3 tests
- ReasoningHard problems that need careful thinking02Qwen3 235B A22B2 of 3 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own01Qwen3 235B A22B1 of 1 test
- Long documentsFinding answers in very long texts11Even1 each
- Following instructionsDoing exactly what it is asked11Even1 each
Agents 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 costs 67% 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-V3 pulls ahead
- Recent programming contest problemsLiveCodeBench v6+9.9points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+9.1points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+7.6points ahead
Where Qwen3 235B A22B pulls ahead
- Keeps track of context across a multi-turn chatMulti-Challenge+9.8points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+8.3points ahead
- Very hard expert questions across many subjectsHumanity's Last Exam+6.3points ahead
Every test, side by side
All 55 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3
- LiveCodeBench v6DeepSeek-V3 by 9.946.937+9.9
- Terminal-Bench HardDeepSeek-V3 by 9.115.26.1+9.1
- SWE-bench VerifiedDeepSeek-V3 by 7.64234.4+7.6
- SciCodetie3939.9tie
ReasoningQwen3 235B A22B
- Humanity's Last ExamQwen3 235B A22B by 6.34.711+6.3
- GPQA DiamondQwen3 235B A22B by 4.565.570+4.5
- CritPttie00tie
FactsDeepSeek-V3
- AA-Omniscience · Non-hallucinationQwen3 235B A22B by 8.314.122.4+8.3
- AA-Omniscience · AccuracyDeepSeek-V3 by 725.418.5+7
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3 by 3.26.19.3+3.2
Long documentsEven
- AA-LCRDeepSeek-V3 by 40.740.70+40.7
- LongBench v2Qwen3 235B A22B by 1.448.750.1+1.4
Following instructionsEven
- Multi-ChallengeQwen3 235B A22B by 9.831.441.2+9.8
- IFBenchDeepSeek-V3 by 2.34138.7+2.3
Other results40 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.
- AIME 2024Qwen3 235B A22B by 46.539.285.7+46.5
- ZebraLogicDeepSeek-V3 by 46.38437.7+46.3
- livecodebench_hardQwen3 235B A22B by 39.814.254+39.8
- livecodebench_mediumQwen3 235B A22B by 35.353.588.8+35.3
- HMMT Feb. 2025Qwen3 235B A22B by 33.329.262.5+33.3
- AIME 2025Qwen3 235B A22B by 30.251.381.5+30.2
- τ²-Bench Telecom (AA run)DeepSeek-V3 by 23.147.124+23.1
- LiveCodeBenchQwen3 235B A22B by 21.549.270.7+21.5
- MultiPL-EDeepSeek-V3 by 17.283.165.9+17.2
- livecodebench_easyQwen3 235B A22B by 15.883.399.1+15.8
- HMMT 2025DeepSeek-V3 by 15.627.511.9+15.6
- OJBenchDeepSeek-V3 by 12.72411.3+12.7
- Tau2 airlineDeepSeek-V3 by 12.53926.5+12.5
- τ²-Bench (Retail)DeepSeek-V3 by 12.169.157+12.1
- SimpleQADeepSeek-V3 by 11.724.913.2+11.7
- τ²-BenchDeepSeek-V3 by 10.432.522.1+10.4
- SuperGPQADeepSeek-V3 by 9.653.744.1+9.6
- MMLU-ProDeepSeek-V3 by 7.775.968.2+7.7
- PolyMath-enDeepSeek-V3 by 7.659.551.9+7.6
- MMMLUQwen3 235B A22B by 7.379.486.7+7.3
- MBPPQwen3 235B A22B by 675.481.4+6
- Artificial Analysis Coding IndexDeepSeek-V3 by 5.62317.4+5.6
- AutoLogiDeepSeek-V3 by 5.688.983.3+5.6
- CNMO 2024Qwen3 235B A22B by 5.443.248.6+5.4
- LiveBenchQwen3 235B A22B by 4.772.477.1+4.7
- Arena HardQwen3 235B A22B by 4.291.495.6+4.2
- AA-OmniscienceDeepSeek-V3 by 4-40.7-44.7+4
- MMLU-ProXDeepSeek-V3 by 3.870.566.7+3.8
- MGSMQwen3 235B A22B by 3.779.883.5+3.7
- vectara_factual_consistencyDeepSeek-V3 by 3.293.990.7+3.2
- IFEvalDeepSeek-V3 by 2.986.183.2+2.9
- vectara_answer_rateDeepSeek-V3 by 2.697.594.9+2.6
- vectara_avg_summary_lengthQwen3 235B A22B by 23.9 rating points81.7105.6+23.9 rating
- AceBenchDeepSeek-V3 by 2.272.770.5+2.2
- MMLU-ReduxDeepSeek-V3 by 1.789.187.4+1.7
- MMLUDeepSeek-V3 by 1.588.587+1.5
- BBHQwen3 235B A22B by 1.487.588.9+1.4
- MATH-500 (EM)Qwen3 235B A22B by 190.291.2+1
- GSM8Ktie9494.4tie
- AA Intelligencetie9.79.5tie
Questions people ask
Which is better, DeepSeek-V3 or Qwen3 235B A22B?
DeepSeek-V3 and Qwen3 235B A22B each win two of the six areas where both have results. DeepSeek-V3 wins coding and facts; Qwen3 235B A22B wins agents and reasoning. DeepSeek-V3 costs 67% less. They are level on long documents and following instructions.
Which is better for coding?
DeepSeek-V3. It wins 3 of the 4 coding tests both models report; Qwen3 235B A22B 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; Qwen3 235B A22B costs $0.70 and $2.80. That makes DeepSeek-V3 about 67% cheaper for the same work.
How do you compare the two?
We use the 55 benchmark tests both models have published scores on. The verdict counts the 15 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 40 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.