DeepSeek permanently retired its legacy API model names on Friday afternoon. As of 15:59 UTC, calls to deepseek-chat and deepseek-reasoner no longer resolve.
The cutover
The retirement time is stated verbatim in DeepSeek's own API documentation. Anything still calling the old names started returning errors during the window rather than degrading gracefully — a hard cutoff at one of the most widely integrated low-cost inference providers.
Not a surprise, and not an announcement
The deprecation was published on April 24, a three-month notice. Several aggregators have framed Friday as the announcement date; it is the execution date. During the grace period the aliases had been transparently routing to deepseek-v4-flash in non-thinking and thinking modes, which is why many integrations kept working without anyone noticing they had already migrated.
What replaces them
Two models: deepseek-v4-pro with 49 billion active parameters, and deepseek-v4-flash with 13 billion active. Total parameter counts circulating in third-party coverage do not appear in DeepSeek's own changelog and should not be attributed to the company.
The migration
Teams pinned to deepseek-chat had been running on v4-flash for weeks without changing a line of code, because the alias routed silently. That is convenient right up until the alias disappears, at which point a change nobody noticed becomes an outage. The fix is a one-line identifier swap; the cost is that nobody knew which model they had been evaluating against.
The wider habit
Model aliases are load-bearing infrastructure that most teams treat as configuration. DeepSeek gave three months' notice, documented the exact minute and published the replacement names — roughly the ceiling of what a provider can do. It still broke integrations, because deprecation notices are read by people and not by production config.