A fixed invoice of 1 million billed input tokens and 100,000 output tokens at peak rates costs $0.4200 on DeepSeek's new model. Make 90% of that input match cache and the same invoice costs $0.1554 — a 63.0% cut, and the output line never moved. That single piece of arithmetic is the design of DeepSeek-V4.1-Flash.
Last issue, the story was the open-model squeeze and the price cuts that came with it. This week the squeeze went structural: DeepSeek shipped V4.1-Flash on September 10 under an MIT license, put the weights on Hugging Face, and made cache pricing the actual weapon. Headline input is $0.15 per million tokens off-peak, output $0.60, cache hits $0.003. Peak doubles every one of those lines at once.
Peak is not a vague surge window. It is 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. Everything else is off-peak, weekends included.
Peak is not a vague surge window. It is 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday.
Convert that to local time and the rate card becomes a map. A Seoul team works a standard 09:00–18:00 day, six of whose nine hours fall inside peak at Korean time. A US Pacific team's entire workday is off-peak. European mornings get hit. The practical read: batch evals, nightly re-indexing, bulk extraction and regression suites are the workloads that can move, and moving them halves the bill. Interactive traffic cannot move and should be budgeted at the peak column.
The architecture is why the price exists. The model runs 40 transformer layers as a 20-layer causal encoder followed by a 20-layer decoder, activating only 8 billion parameters per token during prefill and 16 billion during decode — against a 552B backbone. You are billed for something that behaves like a much smaller model than its parameter count implies.
The KV cache is where that shows up commercially. DeepSeek reports 890 bytes per token globally, roughly a quarter of V4-Flash's footprint, with persistent SSD cache down to about one eighth. Its own chart puts the reduction at 437-fold against V1. For long agent runs that re-read repositories and tool histories, that is the difference between a context that fits and one that does not.
But the sharper number is the spread. Peak cache-hit input is $0.006; peak cache-m