Share promptsSign inCreate team library
The Chat Judge verdicts

Which AI is best at what?

Public benchmark results for 19 tests across 9 fields, charted side by side. Every score links to the leaderboard it came from.

Showing DeepSeek's best-ranked model on each benchmark.

1 benchmarkUpdated 19 Oct 2026Overall workOne number for agentic work across finance, coding, law and tax.
  • Vals Index DeepSeek V4.1 Flash 1017 #8 · 61.26%
See charts 3 benchmarksUpdated 19 Oct 2026LegalDrafting, extraction, legal reasoning and US legal research.
  • Legal Benchmarks leaderboard DeepSeek V4 Pro #11 · 70.2%
  • Legal Research Bench DeepSeek V4.1 Flash 1017 #17 · 45.19%
See charts 3 benchmarksUpdated 21 Oct 2026Health & medicalPhysician-graded health conversations, medical coding and exam questions.
  • HealthBench (independent run) DeepSeek V4.1 Flash (max) #2 · 60.9%
  • MedQA DeepSeek V3.2 (Thinking) #20 · 93.92%
See charts 2 benchmarksUpdated 19 Oct 2026FinanceAnalyst research on SEC filings and long credit agreements.
  • Finance Agent DeepSeek V4.1 Flash #7 · 65.43%
See charts 2 benchmarksUpdated 1 Sep 2026CodingFixing real GitHub issues and solving fresh programming problems.
  • SWE-bench Verified DeepSeek V4 Pro 0813 #2 · 96.4%
  • LiveCodeBench DeepSeek V4 Pro 0813 #13 · 87.53%
See charts 2 benchmarksUpdated 19 Oct 2026Agents & tool useMulti-step tasks an agent completes alone in a terminal.
  • Terminal-Bench 4.0 DeepSeek V4.1 Flash 1017 #17 · 31.82%
  • Terminal-Bench 2.1 DeepSeek V4.1 Flash #18 · 74.53%
See charts 4 benchmarksUpdated 21 Oct 2026Reasoning & knowledgeExpert exams, graduate science, contest maths and broad knowledge.
  • GPQA Diamond DeepSeek V4 Pro 0813 #17 · 92.42%
See charts 1 benchmarkUpdated 20 Oct 2026Chat preferenceWhich answers real people prefer in blind side-by-side votes.
  • LMArena Text Arena DeepSeek-V4.1-Flash-20260910 (Max) #36 · 1475
See charts 1 benchmarkUpdated 9 Jan 2026MultilingualReasoning in languages other than English.
  • MGSM DeepSeek R1 #20 · 92.25%
See charts

ChatJudge does not run these tests. Scores are copied as published by each source, with a link and the date we read them, and only change when we refresh from the source. Different leaderboards use different settings, so compare models within one chart, not across charts. Spotted a mistake? Tell us.