DeepSeek V4 Pro 0813 release and tests

DeepSeek V4 Pro 0813 was released on August 12 despite the model name. Pricing is set at ¥0.025 per million input tokens on cache hit, ¥3 on miss, and ¥6 per million output tokens, with a concurrency limit of 500. Prices are current and will rise. A user tested three tasks: website development with three integrated skills and deployment, imitation of 60 design styles for Bento cards, and a 3D block-breaking game with Three.js. Results were reported as acceptable. Another test generated a looping SVG of a pelican hitting an elephant without thinking mode. - DeepSeek-V4-Pro-0813 released on August 12 - Cache-hit input pricing ¥0.025/million tokens, miss ¥3/million, output ¥6/million - Concurrency limited to 500 (Flash has 2500) - Price note: current rates, increases expected - Test 1: website developed and deployed using three skills - Test 2: 60 Bento card styles generated - Test 3: 3D block-breaking game with Three.js - Separate SVG generation of pelican and elephant loop
Key insights
Pricing structure favors cache hits but imposes higher costs on misses and lower concurrency
DeepSeek V4 Pro 0813 offers a very low ¥0.025 per million tokens for cache hits, but the miss price of ¥3 and output price of ¥6 are higher than comparable Flash tiers. A concurrency cap of 500 versus Flash's 2500 further differentiates access. An explicit note indicates prices will increase from the current launch levels.
Model demonstrates competent multi-skill integration and creative generation
A user tested the model on building and deploying a website requiring three skills, producing 60 Bento-style UI cards, and creating a 3D block-breaking game with Three.js. All tasks completed to an acceptable standard, indicating the model can chain tool calls and follow complex creative prompts effectively.