Claude 3 Haiku vs DeepSeek-V3
Wins 0 of 6 areas
—
Wins 4 of 6 areas
Coding · Reasoning · Long documents · Following instructions
DeepSeek-V3 is the stronger all-rounder.
Scores updated · 34 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 software02DeepSeek-V32 of 2 tests
- ReasoningHard problems that need careful thinking01DeepSeek-V31 of 3 tests · 2 ties
- Long documentsFinding answers in very long texts01DeepSeek-V31 of 1 test
- Following instructionsDoing exactly what it is asked01DeepSeek-V31 of 1 test
- AgentsCarrying out multi-step tasks on its own00Even0 each · 1 tie
- FactsGetting facts right instead of making them up11Even1 each
Agents, long documents and following instructions 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 24% 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 DeepSeek-V3 pulls ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+28.1points ahead
- Code for real scientific research problemsSciCode+20.4points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+14.4points ahead
Where Claude 3 Haiku pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+5.4points ahead
Every test, side by side
All 34 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3
- SciCodeDeepSeek-V3 by 20.418.639+20.4
- Terminal-Bench HardDeepSeek-V3 by 14.40.815.2+14.4
ReasoningDeepSeek-V3
- GPQA DiamondDeepSeek-V3 by 28.137.465.5+28.1
- Humanity's Last Examtie4.14.7tie
- CritPttie00tie
FactsEven
- AA-Omniscience · AccuracyDeepSeek-V3 by 7.917.625.4+7.9
- AA-Omniscience · Non-hallucinationClaude 3 Haiku by 5.419.514.1+5.4
Following instructionsDeepSeek-V3
- IFBenchDeepSeek-V3 by 4.936.141+4.9
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.
- NarrativeQADeepSeek-V3 by 55.224.479.6+55.2
- livecodebench_mediumDeepSeek-V3 by 44.5953.5+44.5
- AIR-Bench 2024Claude 3 Haiku by 41.982.740.8+41.9
- HarmBenchClaude 3 Haiku by 41.691.349.7+41.6
- naturalquestions_closedbookDeepSeek-V3 by 32.314.446.7+32.3
- LiveCodeBenchDeepSeek-V3 by 26.722.549.2+26.7
- τ²-Bench Telecom (AA run)DeepSeek-V3 by 2621.147.1+26
- MMLUDeepSeek-V3 by 22.366.288.5+22.3
- livecodebench_easyDeepSeek-V3 by 22.261.183.3+22.2
- Artificial Analysis Coding IndexDeepSeek-V3 by 16.36.723+16.3
- BBHDeepSeek-V3 by 13.873.787.5+13.8
- DROPDeepSeek-V3 by 13.278.491.6+13.2
- livecodebench_hardDeepSeek-V3 by 12.31.914.2+12.3
- XSTestDeepSeek-V3 by 11.885.397.1+11.8
- OpenBookQADeepSeek-V3 by 11.683.895.4+11.6
- HumanEvalClaude 3 Haiku by 10.775.965.2+10.7
- AA-OmniscienceDeepSeek-V3 by 7.9-48.6-40.7+7.9
- ARC-ChallengeDeepSeek-V3 by 6.189.295.3+6.1
- GSM8KDeepSeek-V3 by 5.188.994+5.1
- MGSMDeepSeek-V3 by 4.775.179.8+4.7
- simple_safety_testsClaude 3 Haiku by 4.710095.3+4.7
- AA IntelligenceDeepSeek-V3 by 4.15.69.7+4.1
- HellaSwagDeepSeek-V3 by 385.988.9+3
- anthropic_red_teamClaude 3 Haiku by 2.910097.1+2.9
Questions people ask
Which is better, Claude 3 Haiku or DeepSeek-V3?
DeepSeek-V3 wins four of the six areas where both have results: coding, reasoning, long documents and following instructions. Claude 3 Haiku wins none. They are level on agents and facts.
Which is better for coding?
DeepSeek-V3. It wins 2 of the 2 coding tests both models report; Claude 3 Haiku wins none.
Which is cheaper?
Claude 3 Haiku costs $0.25 per million input tokens and $1.25 per million output tokens; DeepSeek-V3 costs $0.24 and $0.90. That makes DeepSeek-V3 about 24% cheaper for the same work.
How do you compare the two?
We use the 34 benchmark tests both models have published scores on. The verdict counts the 10 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.