DeepSeek shipped an MIT-licensed model that speaks Codex natively, and it runs a task for three cents.
DeepSeek-V4-Flash-0731 went to public beta on the API yesterday: 284B total parameters, 13B active, a 1M-token context, MIT licence, weights on Hugging Face.1 The line that matters is buried in the changelog — it natively supports the Responses API format and is specifically adapted for Codex. The harness you already run can point at open weights without a translation layer.
Yesterday a bad convention in your repo was a nuisance. Today the arithmetic of which model reads it has changed by two orders of magnitude, and that is a procurement decision, not a preference.
the verbosity objection resolves in its favour. It emits 210M tokens on the index run against a median of 100M, and still costs $0.03 per task where GPT-5.6 Sol costs $1.86. Nine index points, sixty times the money.
Deeper — what migrating actually involves
- The base URL does not change. Set
model=deepseek-v4-flashand the existing OpenAI ChatCompletions or Anthropic-shaped client keeps working; V4-Flash additionally speaks the Responses API format natively, which is what makes the Codex path a configuration change rather than an adapter. - Check your legacy model names first. DeepSeek's changelog lists
deepseek-chatanddeepseek-reasoneras discontinued from 2026-07-24. If either string is still in your code, verify what it resolves to before you assume anything. - The 0731 build keeps the architecture and parameter count of the preview and was only re-post-trained. Anything you benchmarked against the preview has to be re-run, not extrapolated.
- V4-Pro is untouched by this update and its general release will follow soon with no date attached. Do not put it in a plan.
- Where the economics break down: at 1M context and 210M tokens of thinking on a hard evaluation, latency and rate limits become the binding constraint long before cost does. Price is the easy part of this migration.