DeepSeek-V3 vs Mercury 2
Wins 1 of 6 areas
Facts
Wins 5 of 6 areas
Coding · Agents · Reasoning · Long documents · Following instructions
Mercury 2 is the stronger all-rounder.DeepSeek-V3 is better at facts.
Scores updated · 24 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.
- ReasoningHard problems that need careful thinking02Mercury 22 of 3 tests · 1 tie
- CodingWriting and fixing software12Mercury 22 of 3 tests
- AgentsCarrying out multi-step tasks on its own01Mercury 21 of 3 tests · 2 ties
- Long documentsFinding answers in very long texts01Mercury 21 of 1 test
- Following instructionsDoing exactly what it is asked01Mercury 21 of 1 test
- FactsGetting facts right instead of making them up30DeepSeek-V33 of 3 tests
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.
Mercury 2 costs 12% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Mercury 2 pulls ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+28.8points ahead
- Very hard expert questions across many subjectsHumanity's Last Exam+12.4points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+11.5points ahead
Where DeepSeek-V3 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+5.3points ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+4.3points ahead
- Code for real scientific research problemsSciCode+1.3points ahead
Every test, side by side
All 24 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingMercury 2
- Terminal-Bench HardMercury 2 by 11.315.226.5+11.3
- Terminal-Bench 2.1Mercury 2 by 10.416.927.3+10.4
- SciCodeDeepSeek-V3 by 1.33937.7+1.3
AgentsMercury 2
- τ-Bench V3 · BankingMercury 2 by 4.84.79.5+4.8
- GDPValtie00tie
- Terminal-Bench 4.0tie00tie
ReasoningMercury 2
- Humanity's Last ExamMercury 2 by 12.44.717.1+12.4
- GPQA DiamondMercury 2 by 11.565.577+11.5
- CritPttie00.8tie
FactsDeepSeek-V3
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3 by 6.26.112.3+6.2
- AA-Omniscience · Non-hallucinationDeepSeek-V3 by 5.314.18.8+5.3
- AA-Omniscience · AccuracyDeepSeek-V3 by 4.325.421.2+4.3
Following instructionsMercury 2
- IFBenchMercury 2 by 28.84169.8+28.8
Other results10 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 2025Mercury 2 by 39.851.391.1+39.8
- τ²-Bench Telecom (AA run)Mercury 2 by 23.747.170.8+23.7
- LiveCodeBenchMercury 2 by 17.849.267+17.8
- Tau2 airlineMercury 2 by 143953+14
- AA-OmniscienceDeepSeek-V3 by 10-40.7-50.7+10
- Artificial Analysis Coding IndexMercury 2 by 8.12331.1+8.1
- vectara_avg_summary_lengthMercury 2 by 67.4 rating points81.7149.1+67.4 rating
- vectara_factual_consistencyDeepSeek-V3 by 6.293.987.7+6.2
- AA IntelligenceMercury 2 by 4.19.713.8+4.1
- vectara_answer_rateMercury 2 by 2.597.5100+2.5
Questions people ask
Which is better, DeepSeek-V3 or Mercury 2?
Mercury 2 wins five of the six areas where both have results: coding, agents, reasoning, long documents and following instructions. DeepSeek-V3 wins facts.
Which is better for coding?
Mercury 2. It wins 2 of the 3 coding tests both models report; DeepSeek-V3 wins 1.
Which is cheaper?
DeepSeek-V3 costs $0.24 per million input tokens and $0.90 per million output tokens; Mercury 2 costs $0.25 and $0.75. That makes Mercury 2 about 12% cheaper for the same work.
How do you compare the two?
We use the 24 benchmark tests both models have published scores on. The verdict counts the 14 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 10 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.