As recorded at collection: V4.1 Flash rates for DeepSeek V4 Pro requests
As recorded at collection: LLM API service DeepSeek adjusts Flash-series input/output unit prices from September 10, 2026, 12:00 (Beijing time), offering V4.1 Flash at the same price as V4 Flash. Cache-hit input is 0.02 yuan per unit, cache-miss input 1 yuan, output 4 yuan; peak hours are billed at 2× off-peak.
- What you get
- V4.1 Flash at the same price as V4 Flash, with lower Flash-series prices
- Price
- Cached input hits CN¥0.02 (Chinese yuan) per unit, input cache misses CN¥1 per unit, output CN¥4 per unit; peak rates are 2 times off-peak rates
- Valid until
- Price adjustment starts September 10, 2026, 12:00 Beijing time
- Requirements
- Peak rates are 2 times off-peak rates. Once V4.1 Flash launches officially, V4 Pro requests will also be handled by V4.1 Flash and billed at V4.1 Flash rates until V4.1 Pro launches.
Details
DeepSeek is a large language model (LLM) API service billing by input and output units. Flash-series prices change from September 10, 2026, 12:00 Beijing time. Cached input hits cost CN¥0.02 (Chinese yuan) per unit, input cache misses CN¥1, and output CN¥4. Each peak rate is 2 times its off-peak rate.
V4.1 Flash is planned for an official launch around September 10, 2026, after which it will cost the same as V4 Flash. Until V4.1 Pro launches, all requests submitted to V4 Pro will be handled by V4.1 Flash and billed at V4.1 Flash rates.
Users should check whether peak rates apply and plan usage with the difference from off-peak rates in mind. The source asks users to promptly send DeepSeek feedback if problems occur while comparing V4 Pro and V4.1 Flash.
Sources and verification history
- Collection channel
- linux.do
- Collected at
- September 9, 2026, 16:42 KST
- First published
- Updated
- Updates
- September 9, 21:03 · Flash series pricing adjusted, and conditions for applying the V4.1 Flash unit price to V4 Pro requests clarified
- Related posts
- V4.1 Flash rates for DeepSeek V4 Pro requests / V4.1 Flash pricing applied to DeepSeek V4 Pro requests