Model Efficiency
High impactLive published free itemDeepSeek ships V4.1-Flash at $0.15 per million input tokens
MIT-licensed open weights, one million token context, and off-peak input at $0.15 per million. Cached input drops to $0.003 per million.
What changed
DeepSeek released V4.1-Flash on 10 September 2026 under an MIT licence with weights on Hugging Face. Off-peak pricing is $0.15 per million input and $0.60 per million output, rising to $0.30 and $1.20 at peak. Cached input is $0.003 per million off-peak. Context runs to one million tokens. The model is 552B parameters with 8B activated per token read. A reasoning effort dial from 1 to 100 trades accuracy against cost per request. Calls to deepseek-v4-pro route to V4.1-Flash and bill at Flash rates from 14 September.
Why it matters
This is the cheap tier moving, and the cheap tier is where most agent turns actually live. Cached input at $0.003 per million makes repeated-prefix work close to free, which changes the maths on any loop that resends the same system context. The effort dial is unusual: it is a cost control exposed to the caller rather than a model choice.
Take
MIT-licensed open weights, one million token context, and off-peak input at $0.15 per million. Cached input drops to $0.003 per million.
What to do
If you run high-volume classification, extraction or first-draft work on a frontier model, price one week of that workload at Flash off-peak rates before your next billing cycle. Check whether your calls name deepseek-v4-pro, because those reroute and reprice on 14 September whether you act or not.
Evidence
No public sources are linked yet.