DeepSeek-V3 vs GPT-5
Wins 0 of 6 areas
—
Wins 6 of 6 areas
Coding · Agents · Reasoning · Facts · Long documents · Following instructions
GPT-5 is the stronger all-rounder.DeepSeek-V3 is cheaper.
Scores updated · 41 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 software05GPT-55 of 5 tests
- ReasoningHard problems that need careful thinking04GPT-54 of 4 tests
- AgentsCarrying out multi-step tasks on its own02GPT-52 of 2 tests
- Following instructionsDoing exactly what it is asked02GPT-52 of 2 tests
- FactsGetting facts right instead of making them up12GPT-52 of 3 tests
- Long documentsFinding answers in very long texts01GPT-51 of 1 test
Long documents 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 90% 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 GPT-5 pulls ahead
- Recent programming contest problemsLiveCodeBench v6+40.1points ahead
- Keeps track of context across a multi-turn chatMulti-Challenge+38.2points ahead
- Common-sense trick questionsSimpleBench+37.8points ahead
Where DeepSeek-V3 pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 41 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGPT-5
- LiveCodeBench v6GPT-5 by 40.146.987+40.1
- SWE-bench VerifiedGPT-5 by 32.94274.9+32.9
- Terminal-Bench 2.1GPT-5 by 18.316.935.2+18.3
- Terminal-Bench HardGPT-5 by 17.415.232.6+17.4
- SciCodeGPT-5 by 3.93942.9+3.9
ReasoningGPT-5
- SimpleBenchGPT-5 by 37.818.956.7+37.8
- Humanity's Last ExamGPT-5 by 23.84.728.5+23.8
- GPQA DiamondGPT-5 by 19.965.585.4+19.9
- CritPtGPT-5 by 5.705.7+5.7
FactsGPT-5
- AA-Omniscience · AccuracyGPT-5 by 14.825.440.3+14.8
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3 by 96.115.1+9
- AA-Omniscience · Non-hallucinationGPT-5 by 3.714.117.8+3.7
Following instructionsGPT-5
- Multi-ChallengeGPT-5 by 38.231.469.6+38.2
- IFBenchGPT-5 by 32.14173.1+32.1
Other results24 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 2025GPT-5 by 65.827.593.3+65.8
- HMMT Feb. 2025GPT-5 by 59.129.288.3+59.1
- τ²-BenchGPT-5 by 49.932.582.4+49.9
- HarmBenchGPT-5 by 47.949.797.6+47.9
- AIR-Bench 2024GPT-5 by 46.940.887.7+46.9
- AIME 2025GPT-5 by 43.351.394.6+43.3
- Aider-PolyglotGPT-5 by 38.449.688+38.4
- τ²-Bench Telecom (AA run)GPT-5 by 37.747.184.8+37.7
- LiveCodeBenchGPT-5 by 34.349.283.5+34.3
- AA-OmniscienceGPT-5 by 32-40.7-8.7+32
- HumanEvalGPT-5 by 28.265.293.4+28.2
- Tau2 airlineGPT-5 by 23.63962.6+23.6
- Artificial Analysis Coding IndexGPT-5 by 14.82337.8+14.8
- AA IntelligenceGPT-5 by 13.39.723+13.3
- τ²-Bench (Retail)GPT-5 by 1269.181.1+12
- MMLU-ProGPT-5 by 11.275.987.1+11.2
- vectara_factual_consistencyDeepSeek-V3 by 993.984.9+9
- vectara_avg_summary_lengthGPT-5 by 81 rating points81.7162.7+81 rating
- MMLU-ReduxGPT-5 by 6.289.195.3+6.2
- simple_safety_testsGPT-5 by 4.595.399.8+4.5
- vectara_answer_rateGPT-5 by 2.497.599.9+2.4
- anthropic_red_teamGPT-5 by 297.199.1+2
- bbqtie96.796.8tie
- XSTesttie97.197.1tie
Questions people ask
Which is better, DeepSeek-V3 or GPT-5?
GPT-5 wins all six areas where both have results: coding, agents, reasoning, facts, long documents and following instructions. DeepSeek-V3 wins none, but costs 90% less.
Which is better for coding?
GPT-5. It wins 5 of the 5 coding tests both models report; DeepSeek-V3 wins none.
Which is cheaper?
DeepSeek-V3 costs $0.24 per million input tokens and $0.90 per million output tokens; GPT-5 costs $1.25 and $10.00. That makes DeepSeek-V3 about 90% cheaper for the same work.
How do you compare the two?
We use the 41 benchmark tests both models have published scores on. The verdict counts the 17 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 24 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.