DeepSeek-V3 wins more areas, narrowly.GPT-4o is better at long documents.
Scores updated · 72 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 software40DeepSeek-V34 of 4 tests
- ReasoningHard problems that need careful thinking20DeepSeek-V32 of 3 tests · 1 tie
- FactsGetting facts right instead of making them up21DeepSeek-V32 of 3 tests
- Long documentsFinding answers in very long texts01GPT-4o1 of 2 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own00Even0 each · 1 tie
- 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 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 DeepSeek-V3 pulls ahead
- Recent programming contest problemsLiveCodeBench v6+16points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+12.9points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+8.8points ahead
Where GPT-4o pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+48points ahead
- Reasons across sets of long documentsAA-LCR+15.3points ahead
- Keeps track of context across a multi-turn chatMulti-Challenge+8.9points ahead
Every test, side by side
All 72 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3
- LiveCodeBench v6DeepSeek-V3 by 1646.930.9+16
- SWE-bench VerifiedDeepSeek-V3 by 8.84233.2+8.8
- Terminal-Bench HardDeepSeek-V3 by 6.915.28.3+6.9
- SciCodeDeepSeek-V3 by 5.73933.3+5.7
ReasoningDeepSeek-V3
- GPQA DiamondDeepSeek-V3 by 12.965.552.6+12.9
- Humanity's Last ExamDeepSeek-V3 by 2.94.71.8+2.9
- CritPttie00tie
FactsDeepSeek-V3
- AA-Omniscience · Non-hallucinationGPT-4o by 4814.162.1+48
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3 by 3.56.19.6+3.5
- AA-Omniscience · AccuracyDeepSeek-V3 by 1.825.423.7+1.8
Long documentsGPT-4o
- AA-LCRGPT-4o by 15.340.756+15.3
- LongBench v2tie48.748.1tie
Following instructionsEven
- Multi-ChallengeGPT-4o by 8.931.440.3+8.9
- IFBenchDeepSeek-V3 by 54136+5
Other results57 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 2025DeepSeek-V3 by 39.651.311.7+39.6
- CodeforcesDeepSeek-V3 by 375 rating points1134759+375 rating
- Codeforces (Rating)DeepSeek-V3 by 375 rating points1134759+375 rating
- Codeforces (Percentile)DeepSeek-V3 by 35.158.723.6+35.1
- HarmBenchGPT-4o by 33.249.782.9+33.2
- CNMO 2024DeepSeek-V3 by 32.443.210.8+32.4
- AA-OmniscienceGPT-4o by 30.2-40.7-10.5+30.2
- AIME 2024DeepSeek-V3 by 29.939.29.3+29.9
- HumanEvalGPT-4o by 2565.290.2+25
- AIR-Bench 2024GPT-4o by 21.640.862.4+21.6
- livecodebench_mediumDeepSeek-V3 by 21.453.532.1+21.4
- Aider-PolyglotDeepSeek-V3 by 18.949.630.7+18.9
- AlpacaEval2.0 (LC-winrate)DeepSeek-V3 by 18.97051.1+18.9
- τ²-Bench Telecom (AA run)DeepSeek-V3 by 18.247.128.9+18.2
- MATH-500 (EM)DeepSeek-V3 by 15.690.274.6+15.6
- MMLUDeepSeek-V3 by 14.788.573.8+14.7
- SimpleQAGPT-4o by 13.324.938.2+13.3
- Arena HardDeepSeek-V3 by 12.191.479.3+12.1
- SuperGPQADeepSeek-V3 by 11.353.742.4+11.3
- LiveCodeBenchDeepSeek-V3 by 10.949.238.3+10.9
- MGSMGPT-4o by 10.779.890.5+10.7
- C-EvalDeepSeek-V3 by 10.586.576+10.5
- livecodebench_hardDeepSeek-V3 by 9.714.24.5+9.7
- MMLU-ProXDeepSeek-V3 by 9.470.561.1+9.4
- Chinese SimpleQA (C-SimpleQA)DeepSeek-V3 by 9.36858.7+9.3
- DROPDeepSeek-V3 by 8.291.683.4+8.2
- DROP (3-shot F1)DeepSeek-V3 by 7.991.683.7+7.9
- FRAMES (Acc.)GPT-4o by 7.273.380.5+7.2
- Aider-Edit (Acc.)DeepSeek-V3 by 6.879.772.9+6.8
- Tau2 airlineGPT-4o by 6.53945.5+6.5
- τ²-Bench (Retail)DeepSeek-V3 by 5.769.163.4+5.7
- Arena-Hard (GPT-4-1106 judge)DeepSeek-V3 by 5.185.580.4+5.1
- IFEvalDeepSeek-V3 by 5.186.181+5.1
- MATHDeepSeek-V3 by 4.990.285.3+4.9
- vectara_avg_summary_lengthGPT-4o by 4.981.786.6+4.9
- vectara_answer_rateDeepSeek-V3 by 3.797.593.8+3.7
- vectara_factual_consistencyDeepSeek-V3 by 3.593.990.4+3.5
- IFEval (avg)DeepSeek-V3 by 3.287.384.1+3.2
- simple_safety_testsGPT-4o by 3.295.398.5+3.2
- GSM8KDeepSeek-V3 by 3.19490.9+3.1
- CLUEWSCDeepSeek-V3 by 390.987.9+3
- naturalquestions_closedbookGPT-4o by 2.946.749.6+2.9
- MBPP+ (EvalPlus-augmented)DeepSeek-V3 by 2.678.876.2+2.6
- AA IntelligenceDeepSeek-V3 by 2.49.77.3+2.4
- HumanEval-Mul (Pass@1)DeepSeek-V3 by 2.182.680.5+2.1
- anthropic_red_teamGPT-4o by 297.199.1+2
- MMMLUGPT-4o by 279.481.4+2
- bbqDeepSeek-V3 by 1.696.795.1+1.6
- LongBench v2 overall (w/o CoT)GPT-4o by 1.448.750.1+1.4
- OpenBookQAGPT-4o by 1.495.496.8+1.4
- Artificial Analysis Coding IndexGPT-4o by 1.22324.2+1.2
- MMLU-ProDeepSeek-V3 by 1.275.974.7+1.2
- MMLU-ReduxDeepSeek-V3 by 1.189.188+1.1
- livecodebench_easytie83.382.5tie
- DROP (F1)tie8989.2tie
- XSTesttie97.197.3tie
- NarrativeQAtie79.679.5tie
Questions people ask
Which is better, DeepSeek-V3 or GPT-4o?
DeepSeek-V3 wins three of the six areas where both have results: coding, reasoning and facts. GPT-4o wins long documents. They are level on agents and following instructions.
Which is better for coding?
DeepSeek-V3. It wins 4 of the 4 coding tests both models report; GPT-4o wins none.
Which is cheaper?
DeepSeek-V3 costs $0.24 per million input tokens and $0.90 per million output tokens; GPT-4o costs $5.00 and $15.00. That makes DeepSeek-V3 about 94% cheaper for the same work.
How do you compare the two?
We use the 72 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 57 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.