deepseek-chat dies July 24 — and your model already changed under you in April

If deepseek-chat or deepseek-reasoner appears anywhere in your codebase, you have a hard deadline: both names become fully inaccessible after July 24, 2026, 15:59 UTC per DeepSeek's official notice. Two facts to internalize. First, your model already changed — since the V4 preview shipped April 24, the legacy names have silently routed to DeepSeek-V4-Flash, so any evals you ran against V3-era models no longer describe what you're serving. Second, on July 24 the aliases stop resolving and requests hard-fail.

The migration is mechanical but has one structural change: thinking is no longer a model name. deepseek-chat → deepseek-v4-flash; deepseek-reasoner → deepseek-v4-flash plus thinking: {"type": "enabled"} in the request body. Base URL and OpenAI-compatible format are unchanged. Cleanup items: keep stripping reasoning_content from assistant turns in multi-turn flows, revisit hardcoded max_tokens (V4 supports 384K output vs the old 64K cap), and re-test tool calls — legacy reasoner didn't support function calling, V4 does in both modes, which may let you collapse a split reasoning/tool-calling architecture into one call.

The economics are the draw: V4-Flash at $0.14/$0.28 per MTok ($0.0028 on cache hits) with a 1M-token context window — 8x V3.2's 128K, per Artificial Analysis — and both Flash and Pro are MIT-licensed with open weights. There's also an Anthropic-compatible endpoint (api.deepseek.com/anthropic) where Claude SDK traffic maps automatically; pin explicit model names in production so an upstream mapping change can't silently swap your model.

Sources: DeepSeek V4 announcement, DeepSeek pricing page, Developers Digest migration guide, Artificial Analysis