DeepSeek-V3 vs o4-mini
Wins 0 of 6 areas
—
Wins 5 of 6 areas
Coding · Agents · Reasoning · Long documents · Following instructions
o4-mini is the stronger all-rounder.DeepSeek-V3 is cheaper.
Scores updated · 33 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 software02o4-mini2 of 3 tests · 1 tie
- ReasoningHard problems that need careful thinking02o4-mini2 of 3 tests · 1 tie
- Following instructionsDoing exactly what it is asked02o4-mini2 of 2 tests
- AgentsCarrying out multi-step tasks on its own01o4-mini1 of 1 test
- Long documentsFinding answers in very long texts01o4-mini1 of 1 test
- FactsGetting facts right instead of making them up11Even1 each · 1 tie
Agents and long documents rest on a single test each.
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 79% 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 o4-mini pulls ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+27.7points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+26.1points ahead
- Reasons across sets of long documentsAA-LCR+20.3points ahead
Where DeepSeek-V3 pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 33 tests both models report. The winning score is in its model's colour; marks a score checked independently.
Codingo4-mini
- SWE-bench Verifiedo4-mini by 26.14268.1+26.1
- SciCodeo4-mini by 7.53946.5+7.5
- Terminal-Bench Hardtie15.215.2tie
Reasoningo4-mini
- GPQA Diamondo4-mini by 12.965.578.4+12.9
- Humanity's Last Examo4-mini by 11.84.716.5+11.8
- CritPttie00.6tie
FactsEven
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3 by 12.56.118.6+12.5
- AA-Omniscience · Non-hallucinationo4-mini by 5.414.119.5+5.4
- AA-Omniscience · Accuracytie25.424.8tie
Following instructionso4-mini
- IFBencho4-mini by 27.74168.7+27.7
- Multi-Challengeo4-mini by 11.631.443+11.6
Other results20 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.
- HarmBencho4-mini by 47.349.797+47.3
- AIME 2025o4-mini by 41.451.392.7+41.4
- AIR-Bench 2024o4-mini by 37.740.878.5+37.7
- livecodebench_hardo4-mini by 34.714.248.9+34.7
- livecodebench_mediumo4-mini by 32.753.586.2+32.7
- LiveCodeBencho4-mini by 28.249.277.4+28.2
- Aider-Polygloto4-mini by 19.349.668.9+19.3
- livecodebench_easyo4-mini by 15.583.398.8+15.5
- vectara_factual_consistencyDeepSeek-V3 by 12.593.981.4+12.5
- τ²-Bench Telecom (AA run)o4-mini by 8.547.155.6+8.5
- AA Intelligenceo4-mini by 79.716.7+7
- AA-Omniscienceo4-mini by 5-40.7-35.7+5
- simple_safety_testso4-mini by 4.795.3100+4.7
- vectara_avg_summary_lengtho4-mini by 46 rating points81.7127.7+46 rating
- bbqDeepSeek-V3 by 2.796.794+2.7
- Artificial Analysis Coding Indexo4-mini by 2.62325.6+2.6
- vectara_answer_rateo4-mini by 1.797.599.2+1.7
- MMLU-ProXDeepSeek-V3 by 1.270.569.3+1.2
- anthropic_red_teamo4-mini by 1.197.198.2+1.1
- XSTesttie97.197.4tie
Questions people ask
Which is better, DeepSeek-V3 or o4-mini?
o4-mini wins five of the six areas where both have results: coding, agents, reasoning, long documents and following instructions. DeepSeek-V3 wins none, but costs 79% less. They are level on facts.
Which is better for coding?
o4-mini. It wins 2 of the 3 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; o4-mini costs $1.10 and $4.40. That makes DeepSeek-V3 about 79% cheaper for the same work.
How do you compare the two?
We use the 33 benchmark tests both models have published scores on. The verdict counts the 13 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 20 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.