DeepSeek (API)
External APIs
DeepSeek serves V4.1 Flash 1017 under the deepseek-flash name, at $0.15 and $0.60 per million off-peak, with a 1M context and open weights.
Key points : V4.1 Flash 1017 · MIT open weights · Cheapest of the frontier line · Off-peak hours.
Models served : the deepseek-flash name has served V4.1 Flash 1017 since 17 October 2026, at $0.15 input and $0.60 output per million off-peak, double that at peak hours, and $0.003 on a cache hit; deepseek-v4-pro carries V4 Pro 0813, at $0.66 and $1.98 off-peak. The V4 Flash 0731 and Vision Exp names are still accepted and served by the 1017.
Where the data goes
The open platform terms leave users the rights to their inputs and assign them the outputs, explicitly allowing the training of other models by distillation. They publish no retention period for API inputs, however, which forbids presenting the absence of retention as a guarantee.
Retention and training
No retention period for inputs is published. The terms deal mainly with the division of roles: the customer answers requests from its own users, on access as well as on erasure. On the use of inputs for training, the editor makes no written commitment at that address.
When to use it from a sovereign platform
This is the lowest price on the frontier line, and the only model on that line whose weights are published under the MIT licence. For sensitive data the point is therefore to run those weights on-premises rather than call the API, since editor-side retention is undocumented.