DeepSeek keeps V4-Pro on its API, reversing the forced Flash switch
The plan to reroute every V4-Pro call to V4.1-Flash is off — but V4-Flash itself is gone, and the new MIT-licensed Flash runs $0.15/M at 1M context.
Copy markdown
The reprieve: V4-Pro isn't going anywhere
DeepSeek had said that from Sept 14, every `deepseek-v4-pro` request would route to V4.1-Flash at Flash prices. After user pushback it reversed course: V4-Pro keeps serving at unchanged billing until a future V4.1-Pro ships. If you queued an urgent migration, you can stand it down.
But V4-Flash is gone — audit your pinned IDs
The reprieve doesn't cover the older Flash. `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` were retired and now silently route to V4.1-Flash. If you pinned either ID, you're already on the new model — re-check your outputs, latency, and image behavior.
deepseek-flash: 1M context, MIT weights, $0.15/M
V4.1-Flash is a 552B sparse MoE that activates ~8B params on input and 16B on output, with a 1M-token window and up to 384K output tokens. The weights are MIT-licensed and the hosted API runs $0.15 in / $0.60 out per million.
Why it's this cheap: a much smaller KV cache
DeepSeek squeezed the global KV cache to roughly 890 bytes per token — about a quarter of what the prior V4-Flash needed — which is what makes million-token context and self-hosting affordable. MIT weights mean you can also run it on your own boxes, not just the API.
The pitch: reportedly edges Opus 5 and GPT-5.6 Sol
On the DeepSWE coding benchmark, V4.1-Flash reportedly noses ahead of Claude Opus 5 and OpenAI's GPT-5.6 Sol at a fraction of their price. These are vendor-adjacent numbers, so run your own eval before you rewire an agent stack around it.