NEW · API PLANS - 9 models, one key, 250M - 2B tokens a month · from /mo See plans →
Home › Models › DeepSeek › DeepSeek V4 Pro

DeepSeek V4 Pro

DeepSeek V4 Pro is the larger reasoning checkpoint: 1T parameters, a 1M context window, and text plus image input. Published scores on record here are 92.4% GPQA Diamond and 93.5% LiveCodeBench. Reach for it when Flash has already missed a hard math, logic, or competitive-coding problem. It is slower and more expensive per call than Flash, so it is a poor default for routine edits.

Heavy mathFormal logicStep-by-step reasoningCompetitive algorithmsDeep Technical Reasoning & Math
Input$1.74/1M
Output$3.48/1M
Cached read$0.145/1M

Live throughput history

source: status.ebotservers.com
Parameters1T
Context1M
Intelligence index23
TPS avg 24h74.3
TTFB p50365 ms
Uptime100.0%
Published benchmarks
Terminal-Bench 2.1 87.9%
LiveCodeBench 93.5%
GPQA Diamond 92.4%
ProofBench v1.1 54%

Call it


 
 

Other models from DeepSeek