DeepSeek's V4 API goes peak/off-peak — cache reads cost up to 12x
The flat-rate era ends: peak output runs ~4.6x higher, only 7 of 24 hours bill high, and the MIT-licensed weights are the way out.
Copy markdown
The new bill, by the token
DeepSeek-V4-Pro output now runs $1.98/M off-peak and $3.96/M at peak, up from a flat $0.87 (about 4.6x). V4-Flash output climbed from $0.28 to $0.66–$1.32/M, and cache-miss input roughly doubles across the line.
Cache reads take the biggest hit
Cached-input tokens — the cheap reads that agent loops and long-context apps lean on — rose 52% to 1,100%, roughly 12x at the top end. Anything replaying large prompts feels this most.
17 of 24 hours stay at the lower rate
Peak billing hits only ~7 hours a day (01:00–04:00 and 06:00–10:00 UTC — Beijing daytime); the other 17 are half price. Shift batch jobs and eval runs off-peak to roughly halve the bill.
What you're paying for: V4-Pro
The hike lands with V4-Pro's GA — Codeforces ELO 3206, LiveCodeBench 93.5, tool calls inside thinking, and a native OpenAI Responses API with one-click Codex setup, so it drops straight into OpenAI-SDK code.
The hedge: MIT-licensed open weights
Both V4-Pro and V4-Flash weights sit on Hugging Face under MIT, so you can self-host around the surcharge entirely — if you have the hardware (8×H100 minimum for Pro).