DeepSeek V4.1 Flash Cuts Costs With a New MIT License

DeepSeek just did the thing every enterprise AI budget owner both fears and hopes for: it made last month's pricing sheet obsolete. On September 10, DeepSeek released DeepSeek V4.1 Flash, the first model in a new architecture family, and backed it with an MIT license and a price cut that ripples through anyone already running DeepSeek in production.

If you're the person who has to defend a line item in next quarter's AI budget, this is worth five minutes even if you've never touched a DeepSeek API key.

What actually shipped on September 10

DeepSeek V4.1 Flash is a 552-billion-parameter mixture-of-experts model, but it only activates 8 billion parameters to read your input and 16 billion to write its output. It reads images natively alongside text, holds a 1-million-token context window, and was trained from scratch on 45 trillion tokens of mixed text and image data.

The bigger change is under the hood. V4.1 Flash uses a causal encoder-decoder layout: 40 transformer layers split into a 20-layer encoder and a 20-layer decoder, where the decoder pulls its key-value (KV) cache from the encoder's final hidden states instead of generating a fresh one at every decoder layer. The practical result is a per-token KV cache about a quarter the size of the previous V4 Flash model's. A smaller KV cache means DeepSeek can serve more concurrent requests on the same GPU memory, which is exactly why the price dropped.

The pricing swing that matters for budgets

DeepSeek V4.1 Flash lists at $0.30 per million input tokens and $1.20 per million output tokens at peak hours, dropping to $0.15 and $0.60 off-peak. That's already cheap. But the number that should catch a project manager's eye is what happened four days later.

Starting at 04:00 UTC on September 14 (noon in Beijing), DeepSeek began routing every request sent to the legacy model ID deepseek-v4-pro straight to V4.1 Flash, billed at Flash rates. Anyone with deepseek-v4-pro hardcoded into a pipeline saw the same request that used to cost $1.32 per million input tokens and $3.96 per million output tokens at peak suddenly cost $0.30 and $1.20, a cut of roughly 4.4x on input and 3.3x on output, without changing a line of code. DeepSeek says the redirect stays in place until a DeepSeek V4.1 Pro model ships.

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <your-deepseek-api-key>" \
  -d '{
    "model": "deepseek-v4.1-flash",
    "messages": [
      {"role": "user", "content": "Summarize this vendor contract in three bullet points."}
    ],
    "stream": false
  }'

This is the OpenAI-compatible endpoint DeepSeek has used since earlier releases, just pointed at the new model ID. If your integration still calls deepseek-v4-pro, check your billing dashboard before you assume the redirect is free money; confirm the model actually serving your traffic and re-run your cost model against the new rate card rather than the one you budgeted from.

Why the MIT license is the bigger story for PMs

DeepSeek shipped the V4.1 Flash weights on Hugging Face under the MIT license, meaning any team can download, fine-tune, and self-host it commercially with no usage restrictions and no royalty clause to negotiate. For a project manager scoping a pilot, that removes a real blocker: procurement and legal don't need to review a custom model license before a proof of concept can start, and a self-hosted fallback becomes a genuine option if the hosted API's pricing or availability shifts again.

What to do with this Monday morning

If a DeepSeek model is anywhere in your architecture diagram, pull up the actual model ID being billed, not the one you wrote in the design doc six months ago. Then rerun your unit economics: a feature that looked marginal at V4 Pro peak pricing might now clear the bar at Flash rates. And if data residency or vendor risk has been a sticking point in your AI roadmap conversations, the MIT license is worth a line in your next steering committee update, because it changes what "we could just self-host it" actually costs to execute.

The lesson that outlasts this specific price cut: don't bake a single vendor's current pricing into a multi-quarter AI roadmap, because the ground can move under it in four days.