DeepSeek-V4-Pro (Apr 2026)

DeepSeekReleased 24 Apr 2026Open

Compare

151140–163Index score#57 of 229 models

3/ 8Core tests taken3 of 6 areas · plus 12 anchor tests

−26Versus the frontier at releasebest then: GPT-5.5Replaced by DeepSeek-V4-Pro

$0.24/ $0.49Price per million tokensinput / output, list price

1.05MContext windowtokens it reads at once

Where it sitson the frontier

GPT-5 reached this score 9 months earlier.

Test resultsby area · tick = best by any model

Reasoning & knowledgeevidence

GPQA Diamond

91%

Humanity's Last Exam

no result

Mathsevidence

FrontierMath T1-3

45%

Codingevidence

SWE-bench Verified

78%

Terminal-Bench

no result

Agentsevidence

APEX-Agents

no result

Novel problemsno evidence yet

ARC-AGI-2

no result

Long tasksno evidence yet

METR time horizon

no result

Independent results from the Epoch AI benchmarking hub, best reasoning setting. The core tests are shown here; anchor tests also inform the score and can be shown above.

What others sayrank on each leaderboard

The Model Indexindependent tests, our method#57of 219 · 151

LMArena · Creative writingcreative writing#41of 386 · 1446 Epoch Capabilities Indexoverall capability from many benchmarks#26of 239 · 155.4 Artificial Analysis Intelligence Indexoverall capability across ten evaluations#23of 207 · 36 LMArena · Mathsmaths questions in chat#50of 373 · 1466 LMArena · Textoverall chat quality, judged by people#54of 388 · 1458 LMArena · Codingcoding questions in chat#72of 383 · 1500 LMArena · WebDevbuilding web apps and front ends#59of 123 · 1463

Further left is better. Leaderboards measure different things: LMArena is people voting blind between two answers, the others are test scores. Click any row for the source.