Update: DeepSeek legacy names hard-fail today; the rename has a trap
deepseek-chat and -reasoner stop resolving today, no grace alias. The catch: thinking is now a request flag, so a naive rename silently drops reasoning.
Copy markdown
No grace alias — hardcoded names error out
After today's retirement, requests to deepseek-chat and deepseek-reasoner return errors instead of quietly redirecting. Audit env vars, queued jobs, cache keys and fallback routes — any hardcoded name breaks with no clear origin.
The rename trap: thinking is now a flag, not a model
deepseek-v4-flash runs non-thinking by default. Swap deepseek-reasoner → v4-flash without extra_body={"thinking":{"type":"enabled"}} and you silently lose the reasoning you had. Change the model name and the thinking flag in the same commit so they never drift.
What the drop-ins cost
deepseek-v4-flash is $0.14 in / $0.28 out per 1M tokens, with cache-hit input near-free (~$0.003/1M); the base URL stays api.deepseek.com. Reasoning-heavy jobs step up to v4-pro (~$0.44 in / $0.87 out, currently discounted) with thinking on.
V4 speaks both OpenAI and Anthropic protocols
The V4 models are exposed through both the OpenAI ChatCompletions interface and an Anthropic-style Messages interface, so you can point either SDK at api.deepseek.com — useful if your stack is already Claude-shaped.