DeepSeek models and APIs

About DeepSeek

DeepSeek's three chat families sit at the low-cost end of the text catalogue while keeping a million-token context window. They are the usual first stop for high-volume workloads where the input dominates the bill and the reply is short.

DeepSeek V4 Pro takes up to 1,048,576 tokens in and writes back up to 384,000 — the largest reply ceiling in the catalogue — and runs at high reasoning effort by default, with low available per request. V4.1 Flash keeps the same window and the same 384,000-token ceiling at a lower rate. V4 Flash trades the long reply for a 65,536-token ceiling and the cheapest output rate of the three. All three accept a reasoning-effort field, so one family covers both a quick classification and a long chain of thought.

Start with V4 Flash for short, frequent turns and move to V4 Pro only when a single reply genuinely has to run past 65,536 tokens. Because a long conversation resends its prefix every turn, cached input is what actually decides the bill here, and it rewards sending a byte-identical prefix rather than rebuilding one.