News & Commentary

DeepSeek model string stops meaning what it says.

DeepSeek reroutes deepseek-v4-pro to its cheaper V4.1-Flash on Sept 14. For bring-your-own-key apps, a silent swap behind a model name is the whole problem.

DeepSeek released V4.1-Flash on Sept 10 - MIT-licensed weights, a million-token context - and announced that from noon Beijing time on Monday Sept 14, API calls naming deepseek-v4-pro will route to Flash and bill at the Flash price, roughly a quarter of what V4 Pro cost. DeepSeek says Flash outperforms the older Pro.

Be fair, because the good parts are real: it is cheaper for the user, and MIT-licensed weights at this scale is a genuine gift - no territory clause, no acceptable-use annex, no revocation. Praise that without pretending 552B parameters fit on a 5090.

Ignore the price war and watch the reroute, because it is being reported as a footnote and it is the story. On Monday, a request naming deepseek-v4-pro stops being served by V4 Pro. Same endpoint, same model string, different model underneath. That breaks the core promise of the bring-your-own-key category I ship into - that the user is talking to the model they chose, on their own account, with no intermediary quietly deciding for them. It is cheaper, which is exactly why nobody objects, and exactly why it becomes normal.

The deadline is Monday. After that, 'which model answered me' is a question a BYO-key app has to be able to answer, not assume.