DeepSeek V4: 1M-context coding from $0.14/M, legacy endpoints retired
The API that undercuts closed frontier models now caches near-free, retires its old aliases, and has AI's first peak-hour surcharge waiting in the wings.
Copy markdown
The rate card, in two numbers
V4-Flash runs $0.14 in / $0.28 out per million tokens; V4-Pro, the reasoning tier, is $0.435 / $0.87 — both with a 1M-token context window. That undercuts every closed frontier API for chat, extraction, and coding work.
deepseek-chat and deepseek-reasoner are gone
The legacy aliases were retired at GA on July 24; requests must now name deepseek-v4-flash or deepseek-v4-pro explicitly. If your code still points at the old endpoints, it's already failing — swap the model IDs today.
Prefix caching cuts input up to 95%
Automatic prefix caching drops cache-hit input to $0.0028 per million — roughly a 50x discount on stable system prompts, with no code changes. Long, fixed contexts effectively read for free.
AI's first surge pricing is coming
DeepSeek has signaled a peak-hour surcharge — reportedly a 2x multiplier in set daily windows — but it isn't live yet, and no rate or start date sits on the official card. Start logging request timestamps now so the eventual bill doesn't blindside you.
What to do this week
Pin explicit model IDs, add timestamp logging for future surge attribution, and benchmark V4-Pro against your current coding model while pricing is still flat and the discount window is wide open.