DeepSeek retired its old model names — migrate to V4 or calls fail
The deepseek-chat and deepseek-reasoner aliases stopped resolving July 24; the naive rename silently turns on reasoning and inflates your token bill.
Copy markdown
The aliases stopped resolving July 24
After 15:59 UTC on July 24, any request with model `deepseek-chat` or `deepseek-reasoner` fails — the names no longer exist, with no grace period and no automatic routing. If a pipeline still hard-codes either one, it is broken right now, not later.
The one-line rename quietly inflates your bill
The drop-in name is `deepseek-v4-flash`, but it defaults thinking mode ON. Rename on autopilot and every call now reasons first — a 200-token reply can balloon past 1,000 output tokens at $0.28/M. For a true `deepseek-chat` swap turn thinking off; to match `deepseek-reasoner`, leave it on.
What replaces them: two V4 tiers, 1M context, still cheap
V4-Flash (284B total / 13B active) runs $0.14 in / $0.28 out per 1M tokens, with cache hits at $0.0028. V4-Pro (1.6T / 49B active) is $0.435 / $0.87. Both carry a 1M-token window and up to 384K output, powered by DeepSeek Sparse Attention.
Reasoner users are NOT auto-upgraded to Pro
Renaming to Flash keeps you on the cheaper tier. To get DeepSeek's frontier model — which the lab says beats every current open model on Math, STEM and coding, and is reported around 80% on SWE-bench Verified — you must explicitly request `deepseek-v4-pro`.
It still speaks both OpenAI and Anthropic formats
V4 serves the OpenAI ChatCompletions and Anthropic Messages APIs, plus tool calls and JSON mode. Existing GPT- or Claude-style clients can point at it with just a base-URL and model-name change — the migration is a config edit, not a rewrite.