DeepSeek-V4-Pro vs Devstral 2
Wins 5 of 6 areas
Coding · Agents · Reasoning · Long documents · Following instructions
Wins 0 of 6 areas
—
DeepSeek-V4-Pro is the stronger all-rounder.
Scores updated · 22 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 software60DeepSeek-V4-Pro6 of 6 tests
- AgentsCarrying out multi-step tasks on its own30DeepSeek-V4-Pro3 of 3 tests
- ReasoningHard problems that need careful thinking30DeepSeek-V4-Pro3 of 3 tests
- Long documentsFinding answers in very long texts10DeepSeek-V4-Pro1 of 1 test
- Following instructionsDoing exactly what it is asked10DeepSeek-V4-Pro1 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.
DeepSeek-V4-Pro costs 46% 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-V4-Pro pulls ahead
- Reasons across sets of long documentsAA-LCR+42.4points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+38.4points ahead
- Very hard expert questions across many subjectsHumanity's Last Exam+33.9points ahead
Where Devstral 2 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+8.8points ahead
Every test, side by side
All 22 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V4-Pro
- Terminal-Bench 2.1DeepSeek-V4-Pro by 33.76430.3+33.7
- Terminal-Bench HardDeepSeek-V4-Pro by 27.346.218.9+27.3
- LMArena · WebDevDeepSeek-V4-Pro by 269 rating points14631194+269 rating
- SWE-bench VerifiedDeepSeek-V4-Pro by 26.880.653.8+26.8
- SciCodeDeepSeek-V4-Pro by 1850.832.8+18
- SWE-bench MultilingualDeepSeek-V4-Pro by 14.976.261.3+14.9
AgentsDeepSeek-V4-Pro
- GDPValDeepSeek-V4-Pro by 30.532.92.4+30.5
- τ-Bench V3 · BankingDeepSeek-V4-Pro by 19.630.110.5+19.6
- Terminal-Bench 4.0DeepSeek-V4-Pro by 14.614.60+14.6
ReasoningDeepSeek-V4-Pro
- Humanity's Last ExamDeepSeek-V4-Pro by 33.937.53.6+33.9
- GPQA DiamondDeepSeek-V4-Pro by 29.488.859.4+29.4
- CritPtDeepSeek-V4-Pro by 12.912.90+12.9
FactsEven
- AA-Omniscience · AccuracyDeepSeek-V4-Pro by 22.24320.8+22.2
- AA-Omniscience · Non-hallucinationDevstral 2 by 8.85.914.7+8.8
Following instructionsDeepSeek-V4-Pro
- IFBenchDeepSeek-V4-Pro by 38.476.538.1+38.4
Other results6 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.
- τ²-Bench Telecom (AA run)DeepSeek-V4-Pro by 71.396.224.9+71.3
- AA-OmniscienceDeepSeek-V4-Pro by 36-10.7-46.7+36
- Terminal-Bench 2.0DeepSeek-V4-Pro by 35.367.932.6+35.3
- Artificial Analysis Coding IndexDeepSeek-V4-Pro by 28.159.431.3+28.1
- AA Agentic IndexDeepSeek-V4-Pro by 22.827.74.9+22.8
- AA IntelligenceDeepSeek-V4-Pro by 21.830.48.6+21.8
Questions people ask
Which is better, DeepSeek-V4-Pro or Devstral 2?
DeepSeek-V4-Pro wins five of the six areas where both have results: coding, agents, reasoning, long documents and following instructions. Devstral 2 wins none. They are level on facts.
Which is better for coding?
DeepSeek-V4-Pro. It wins 6 of the 6 coding tests both models report; Devstral 2 wins none.
Which is cheaper?
DeepSeek-V4-Pro costs $0.43 per million input tokens and $0.87 per million output tokens; Devstral 2 costs $0.40 and $2.00. That makes DeepSeek-V4-Pro about 46% cheaper for the same work.
How do you compare the two?
We use the 22 benchmark tests both models have published scores on. The verdict counts the 16 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 6 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.