Claude 3.5 Haiku vs DeepSeek-V3
Wins 2 of 6 areas
Agents · Following instructions
Wins 3 of 6 areas
Coding · Reasoning · Long documents
DeepSeek-V3 wins more areas, narrowly.Claude 3.5 Haiku is better at agents.
Scores updated · 29 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 software04DeepSeek-V34 of 4 tests
- ReasoningHard problems that need careful thinking02DeepSeek-V32 of 3 tests · 1 tie
- Long documentsFinding answers in very long texts01DeepSeek-V31 of 1 test
- AgentsCarrying out multi-step tasks on its own10Claude 3.5 Haiku1 of 2 tests · 1 tie
- Following instructionsDoing exactly what it is asked10Claude 3.5 Haiku1 of 1 test
- FactsGetting facts right instead of making them up11Even1 each
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.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where DeepSeek-V3 pulls ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+24.7points ahead
- Reasons across sets of long documentsAA-LCR+13.4points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+12.9points ahead
Where Claude 3.5 Haiku pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+44.8points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+1.8points ahead
- Customer-service tasks in a simulated bankτ-Bench V3 · Banking+1.1points ahead
Every test, side by side
All 29 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3
- Terminal-Bench HardDeepSeek-V3 by 12.92.315.2+12.9
- SciCodeDeepSeek-V3 by 11.627.439+11.6
- Terminal-Bench 2.1DeepSeek-V3 by 6.810.116.9+6.8
- SWE-bench VerifiedDeepSeek-V3 by 1.440.642+1.4
AgentsClaude 3.5 Haiku
- τ-Bench V3 · BankingClaude 3.5 Haiku by 1.15.84.7+1.1
- GDPValtie00tie
ReasoningDeepSeek-V3
- GPQA DiamondDeepSeek-V3 by 24.740.865.5+24.7
- Humanity's Last ExamDeepSeek-V3 by 1.13.64.7+1.1
- CritPttie00tie
FactsEven
- AA-Omniscience · Non-hallucinationClaude 3.5 Haiku by 44.858.914.1+44.8
- AA-Omniscience · AccuracyDeepSeek-V3 by 12.313.225.4+12.3
Following instructionsClaude 3.5 Haiku
- IFBenchClaude 3.5 Haiku by 1.842.841+1.8
Other results16 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.
- HumanEvalClaude 3.5 Haiku by 22.988.165.2+22.9
- τ²-Bench Telecom (AA run)DeepSeek-V3 by 22.524.647.1+22.5
- Aider-PolyglotDeepSeek-V3 by 21.62849.6+21.6
- MMLUDeepSeek-V3 by 21.467.188.5+21.4
- AA-OmniscienceClaude 3.5 Haiku by 18.2-22.5-40.7+18.2
- GSM8KDeepSeek-V3 by 12.581.594+12.5
- naturalquestions_closedbookDeepSeek-V3 by 12.334.446.7+12.3
- MMLU-ProDeepSeek-V3 by 10.96575.9+10.9
- MBPPClaude 3.5 Haiku by 10.285.675.4+10.2
- OpenBookQADeepSeek-V3 by 1085.495.4+10
- DROPDeepSeek-V3 by 8.583.191.6+8.5
- Artificial Analysis Coding IndexDeepSeek-V3 by 7.115.923+7.1
- MGSMClaude 3.5 Haiku by 5.885.679.8+5.8
- NarrativeQADeepSeek-V3 by 3.376.379.6+3.3
- MATHDeepSeek-V3 by 387.290.2+3
- AA Intelligencetie8.99.7tie
Questions people ask
Which is better, Claude 3.5 Haiku or DeepSeek-V3?
DeepSeek-V3 wins three of the six areas where both have results: coding, reasoning and long documents. Claude 3.5 Haiku wins agents and following instructions. They are level on facts.
Which is better for coding?
DeepSeek-V3. It wins 4 of the 4 coding tests both models report; Claude 3.5 Haiku wins none.
How do you compare the two?
We use the 29 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 16 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.