DeepSeek API documentation showing V4 model names and deprecation notice Image: DeepSeek / api-docs.deepseek.com
by VibecodedThis

DeepSeek Retires Legacy API Names Today: Migrate from deepseek-chat and deepseek-reasoner Now

DeepSeek's legacy model aliases deepseek-chat and deepseek-reasoner stop working today at 15:59 UTC. Any code still using those names will start returning errors.

Share

DeepSeek is retiring the deepseek-chat and deepseek-reasoner model aliases today. The cutoff is 15:59 UTC, July 24, 2026. After that timestamp, any API call using those names returns an error with no fallback and no grace period.

The deprecation was announced when DeepSeek shipped V4 in April. Both aliases have been routing to V4-Flash during the transition window, so most code should have continued working. That routing ends today.

What to Replace Them With

The migration is a single-line change in most codebases:

  • deepseek-chat (the non-thinking model) becomes deepseek-v4-flash
  • deepseek-reasoner (the thinking model) becomes deepseek-v4-flash with thinking mode enabled, or deepseek-v4-pro if you want higher capability

The base URL stays the same. DeepSeek’s API supports both the OpenAI ChatCompletions format and the Anthropic API format, so the interface itself hasn’t changed.

If you were using deepseek-reasoner for extended reasoning tasks, deepseek-v4-pro is likely the better long-term choice. It has 1.6T total parameters with 49B active per token versus V4-Flash’s 284B/13B split, and both support a 1M-token context window.

What Breaks

Anything that hardcodes the old model names. That includes:

  • Agent definitions and system prompts that specify a model
  • Automated pipelines, scheduled reports, background jobs
  • Infrastructure-as-code and deployment configs
  • Claude Code and OpenCode integrations that use DeepSeek as a backend

The full list of V4 model IDs and current pricing is in the DeepSeek API docs.

Context

DeepSeek V4 shipped April 24 as a preview and introduced two distinct models where V3 had one. The Flash tier targets speed and cost at $0.44 per million output tokens, while Pro targets accuracy at a higher rate. The legacy deepseek-chat and deepseek-reasoner names predate that split and have been a source of confusion ever since.

The July 24 deadline removes any ambiguity. If you’re still on the old names, check your code now.

Share