deepseek-chat and deepseek-reasoner retire on July 24 — check your model IDs now

If your codebase still calls deepseek-chat or deepseek-reasoner, you have until 2026-07-24 15:59 UTC before those aliases stop meaning what you think they mean. After the cutoff, both legacy IDs route to deepseek-v4-flash.

For most chat workloads this is a silent upgrade: Flash is cheaper than the old V3-era pricing and supports a 1M-token context. The risk sits with deepseek-reasoner users. If your pipeline depends on long chain-of-thought output, being silently routed to Flash's default non-thinking mode can change results without an error message. Move those calls to deepseek-v4-pro, or set Flash's thinking mode explicitly.

The fix is usually a one-line change: keep your base URL, swap the model string. Our migration checklist covers the ID mapping table, SDK snippets, and the three failure modes we'd test for before the deadline.