DeepSeek is cutting Flash model prices from September 10
Off-peak input cache hits drop to $0.003 per million tokens from September 10. Peak rates stay double the off-peak rate.
From September 10, a cached Flash input token costs 50x less than an uncached one: $0.003 vs $0.15 per 1M.
DeepSeek is cutting prices on its Flash series of models from September 10, 2026. A notice on the DeepSeek platform sets the new off-peak price of input cache hits at $0.003 per million tokens, input cache misses at $0.15, and output at $0.60. Peak-hour prices will be double the off-peak rates, and the peak windows do not change.
The new numbers
All prices are per million tokens, in US dollars. The change takes effect at 12:00 Beijing time on September 10, which is 04:00 UTC and 09:30 IST. Compared with the rates on DeepSeek's models and pricing page today, the new off-peak prices cut input cache hits by about 57%, input cache misses by 32%, and output by 9%.
- Input cache hit: $0.003 off-peak, $0.006 peak.
- Input cache miss: $0.15 off-peak, $0.30 peak.
- Output: $0.60 off-peak, $1.20 peak.
For reference, the off-peak rates listed on the pricing page today are $0.007 for cache hits, $0.22 for cache misses, and $0.66 for output. The cut is concentrated on input, which is what long agent conversations consume in bulk.
Peak windows stay the same
The structure does not change. Peak hours remain Monday to Friday, 01:00 to 04:00 and 06:00 to 10:00 UTC, which is 09:00 to 12:00 and 14:00 to 18:00 Beijing time, or 06:30 to 09:30 and 11:30 to 15:30 IST. Every other hour, and the whole weekend, is off-peak. Peak prices are double off-peak prices, which is the same deal as an off-peak discount of 50%.
The effective moment is easy to time in one zone: 12:00 Beijing time on September 10 is 09:30 IST, the exact end of that morning's first peak window.
Cache hits are the lever that matters
The three input prices exist because DeepSeek, like most model providers, reuses work across requests. When your call starts with a prompt prefix the provider has already processed, those tokens are billed as cache hits. New text, changed prefixes, and long fresh contexts are cache misses. Output tokens carry their own rate.
At the new prices the gap is stark. One million cached input tokens cost $0.003 off-peak. The same million as a cache miss costs $0.15, 50 times more. Agent workloads resend the same system prompt, tool definitions, and conversation history on every turn, so keeping that prefix stable is what moves the bill onto the cheap tier.
Two practical notes. A prefix only stays cached if the earlier part of the conversation is not rewritten mid-run. And cache behaviour depends on request shape, so the saving shows up in your own traffic, not in the price sheet.
Why a build studio cares
Model API prices are a direct cost input for the AI workflows and agents we build and run, and DeepSeek's Flash tier is one of the cheap endpoints that workhorse agent loops run on. A cut of this size on input changes the monthly math for token-heavy builds, which is why the token bill is modelled as part of the architecture, not as an afterthought.
The honest caveat: the source is a platform notice, and provider pricing has moved several times this year. The windows and the double-peak structure survived those moves, but re-check the pricing page before you lock a cost estimate into a client quote.
Next step: read DeepSeek's models and pricing page and its context caching guide. If you are building an agent workflow and want the token bill modelled before you commit to a provider, write to us at hello@gattyworks.com.