DeepSeek V3.2 vs DeepSeek V4-Pro

Best cell per row highlighted. Null means undisclosed — never counted as zero.

DeepSeek
DeepSeek V3.2GA
8.2
AI Panel
DeepSeek
DeepSeek V4-Propreview
8.5
AI Panel

Identity & lifecycle

ProviderDeepSeekDeepSeek
Family / tierDeepSeek V3 · LargeDeepSeek V4 · Pro
StatusGApreview
Released2025-11-302026-04-23
Knowledge cutoff2025-072026-02

Architecture & context

Context window128K1M
Max output tokens64K384K
Input modalitiestexttext
Reasoning modeoptionaloptional
Open weightsYesYes

Pricing (per Mtok)

Input$0.252$0.435
Output$0.378$0.87
Cached input$0.025$0.0036
Batch input
Free tierYesYes

Speed

Output speed (tok/s)
Time to first token (s)
Latency tiermediummedium

Trust & deployment

Trains on inputsYesYes
LicenseMITMIT
Clouds
Compliance

AI Panel scoring

Unified score8.28.5
Decision Maker88
Domain Strategist88.5
Finance Lead99.5
Domain Practitioner8.58.5
Power User7.57.5
Skeptic7.57.5
Value score9.39.7

Benchmarks

MMLU-Pro8587.5
GPQA Diamond90.1
AIME 202593.1
SWE-bench Verified67.880.6
HumanEval76.8
LiveCodeBench74.193.5
MRCR Long Context83.5
SimpleQA57.9
Humanity's Last Exam30.637.7
LMArena Coding Elo1287
Artificial Analysis Index52

Scores link to their sources. Missing cells mean the vendor hasn't disclosed a result — honesty over padding.