Claude Sonnet 4.6 vs DeepSeek-V4-Pro
Wins 3 of 7 areas
Coding · Agents · Long documents
Wins 4 of 7 areas
Reasoning · Facts · Math · Following instructions
DeepSeek-V4-Pro wins more areas, narrowly.Claude Sonnet 4.6 is better at coding.
Scores updated · 53 tests both models report · How we compare
Where each one wins
Tests won in each of the seven 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 thinking13DeepSeek-V4-Pro3 of 4 tests
- FactsGetting facts right instead of making them up13DeepSeek-V4-Pro3 of 4 tests
- MathCompetition and research-level math02DeepSeek-V4-Pro2 of 2 tests
- Following instructionsDoing exactly what it is asked01DeepSeek-V4-Pro1 of 2 tests · 1 tie
- CodingWriting and fixing software51Claude Sonnet 4.65 of 8 tests · 2 ties
- AgentsCarrying out multi-step tasks on its own43Claude Sonnet 4.64 of 7 tests
- Long documentsFinding answers in very long texts10Claude Sonnet 4.61 of 1 test
Long documents rests on a single test.
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 93% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where DeepSeek-V4-Pro pulls ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+19.9points ahead
- Short factual questions, answered correctlySimpleQA Verified+13.4points ahead
- Multi-step tasks using many real software toolsMCP Atlas+12.3points ahead
Where Claude Sonnet 4.6 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+45.7points ahead
- Programming problems, refreshed regularlyLiveBench · Coding+9.3points ahead
- Command-line tasks in a real terminalTerminal-Bench 2.1+7.2points ahead
Every test, side by side
All 53 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingClaude Sonnet 4.6
- LiveBench · CodingClaude Sonnet 4.6 by 9.379.370+9.3
- Terminal-Bench 2.1Claude Sonnet 4.6 by 7.271.264+7.2
- Terminal-Bench HardClaude Sonnet 4.6 by 6.85346.2+6.8
- LMArena · WebDevClaude Sonnet 4.6 by 58 rating points15221464+58 rating
- SWE-bench ProClaude Sonnet 4.6 by 2.758.155.4+2.7
- SWE-bench VerifiedDeepSeek-V4-Pro by 179.680.6+1
- SciCodetie50.150.8tie
- LiveBench · Agentic Codingtie42.642.6tie
AgentsClaude Sonnet 4.6
- MCP AtlasDeepSeek-V4-Pro by 12.361.373.6+12.3
- Terminal-Bench 4.0DeepSeek-V4-Pro by 11.6314.6+11.6
- BrowseCompDeepSeek-V4-Pro by 8.774.783.4+8.7
- τ-Bench V3 · BankingClaude Sonnet 4.6 by 4.334.430.1+4.3
- GDPValClaude Sonnet 4.6 by 3.836.732.9+3.8
- AA ApexAgentsClaude Sonnet 4.6 by 3.72824.3+3.7
- AA IT-Bench SREClaude Sonnet 4.6 by 1.539.838.3+1.5
ReasoningDeepSeek-V4-Pro
- CritPtDeepSeek-V4-Pro by 9.83.112.9+9.8
- Humanity's Last ExamDeepSeek-V4-Pro by 3.933.637.5+3.9
- LiveBench · ReasoningClaude Sonnet 4.6 by 2.184.882.7+2.1
- GPQA DiamondDeepSeek-V4-Pro by 1.387.588.8+1.3
FactsDeepSeek-V4-Pro
- AA-Omniscience · Non-hallucinationClaude Sonnet 4.6 by 45.751.65.9+45.7
- SimpleQA VerifiedDeepSeek-V4-Pro by 13.432.846.2+13.4
- AA-Omniscience · AccuracyDeepSeek-V4-Pro by 2.140.943+2.1
- Vectara HHEM hallucination ratelower is betterDeepSeek-V4-Pro by 210.68.6+2
MathDeepSeek-V4-Pro
- USAMO 2026DeepSeek-V4-Pro by 5.75560.7+5.7
- LiveBench · MathematicsDeepSeek-V4-Pro by 3.78790.7+3.7
Following instructionsDeepSeek-V4-Pro
- IFBenchDeepSeek-V4-Pro by 19.956.676.5+19.9
- LiveBench · Instruction Followingtie63.262.4tie
Other results25 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.
- AA-OmniscienceClaude Sonnet 4.6 by 22.912.2-10.7+22.9
- τ²-Bench Telecom (AA run)DeepSeek-V4-Pro by 20.575.796.2+20.5
- GDPval-AA (Elo)Claude Sonnet 4.6 by 122 rating points16761554+122 rating
- ToolathlonDeepSeek-V4-Pro by 10.84151.8+10.8
- GDPval-AA v2Claude Sonnet 4.6 by 89 rating points13951306+89 rating
- Terminal-Bench 2.0DeepSeek-V4-Pro by 8.859.167.9+8.8
- AgentWorldBench - MCPClaude Sonnet 4.6 by 6.77063.3+6.7
- AgentWorldBench - TerminalClaude Sonnet 4.6 by 5.75751.3+5.7
- AA Agentic IndexClaude Sonnet 4.6 by 5.433.127.7+5.4
- AgentWorldBench - SWEClaude Sonnet 4.6 by 5.164.559.4+5.1
- vectara_avg_summary_lengthDeepSeek-V4-Pro by 39.1 rating points114.7153.8+39.1 rating
- Artificial Analysis Coding IndexClaude Sonnet 4.6 by 3.66359.4+3.6
- livebench_data_analysisClaude Sonnet 4.6 by 3.57874.5+3.5
- AgentWorldBench - OverallClaude Sonnet 4.6 by 35653+3
- AgentWorldBench - AndroidClaude Sonnet 4.6 by 2.85855.2+2.8
- vectara_answer_rateClaude Sonnet 4.6 by 2.799.997.2+2.7
- GPQA (unspecified)Claude Sonnet 4.6 by 2.189.987.8+2.1
- livebench_languageDeepSeek-V4-Pro by 276.178.1+2
- vectara_factual_consistencyDeepSeek-V4-Pro by 289.491.4+2
- LiveBenchClaude Sonnet 4.6 by 1.975.573.6+1.9
- HLE (with tools)DeepSeek-V4-Pro by 1.446.848.2+1.4
- AgentWorldBench - SearchClaude Sonnet 4.6 by 1.228.827.6+1.2
- AgentWorldBench - OStie63.263.7tie
- AgentWorldBench - Webtie50.850.3tie
- AA Intelligencetie30.130.4tie
Questions people ask
Which is better, Claude Sonnet 4.6 or DeepSeek-V4-Pro?
DeepSeek-V4-Pro wins four of the seven areas where both have results: reasoning, facts, math and following instructions. Claude Sonnet 4.6 wins coding, agents and long documents.
Which is better for coding?
Claude Sonnet 4.6. It wins 5 of the 8 coding tests both models report; DeepSeek-V4-Pro wins 1, and 2 are ties.
Which is cheaper?
Claude Sonnet 4.6 costs $3.00 per million input tokens and $15.00 per million output tokens; DeepSeek-V4-Pro costs $0.43 and $0.87. That makes DeepSeek-V4-Pro about 93% cheaper for the same work.
How do you compare the two?
We use the 53 benchmark tests both models have published scores on. The verdict counts the 28 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 25 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.