DeepSeek
Prices halve outside peak hours, and cache hits land at $0.003 per million input tokens, two orders of magnitude below the miss rate. The naming inverts the usual intuition: the cheaper model is the one that reasons.
Pricing that changes with the clock
Most providers charge the same rate whenever you call. DeepSeek halves its prices outside peak hours, and for a workload you control the timing of, that is the reason to look.
| Model | Context | Cache miss in | Cache hit in | Output |
|---|---|---|---|---|
deepseek-flash | 1M | $0.15 / $0.3 | $0.003 / $0.006 | $0.6 / $1.2 |
deepseek-v4-pro | 1M | $0.66 / $1.32 | $0.022 / $0.044 | $1.98 / $3.96 |
Off-peak rate first, peak second. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays.
Two things follow. A batch job you can schedule outside those windows costs half. And the windows are in UTC on a Chinese holiday calendar, so "off-peak" is not where a European or American team's intuition puts it. Check the clock rather than assuming overnight is cheap.
Cache hits are the real number
Look at the cache-hit column again. Input at $0.003 per million tokens
off-peak on deepseek-flash is two orders of magnitude below the miss rate.
That changes what prompt design is worth. A long stable system prompt or a large retrieved corpus, sent repeatedly, costs almost nothing once cached, which makes the discipline prompt caching describes more valuable here than on providers with a 10% cache rate.
It also means a cost model built on average input tokens will be wrong by a wide margin in either direction depending on your hit rate. Record the cache fields separately, as cost and latency monitoring argues, or you will not know which regime you are in.
Two models, split by thinking
deepseek-flash supports both non-thinking and thinking modes, with thinking the
default. deepseek-v4-pro does not support thinking at all.
That inverts the usual naming intuition, where the more expensive model is the one that reasons. Here the cheaper model is the reasoning one, and the pricier model is the non-thinking option. Read the table rather than the names.
Both carry a 1M-token context window, so the choice between them is not about how much you can send.
What to check before committing
Three things this page cannot settle for you.
Where the data goes. A hosted API means your prompts reach a third party, and which jurisdiction that is matters to some teams and not others. Security and privacy concerns covers establishing retention terms and who the processor is, and that work is the same regardless of provider.
Whether quality holds on your task. Price is the headline and not the decision. Choosing the right model argues for twenty cases from your own workload, and a provider this much cheaper is exactly where the temptation to skip that step is strongest.
Whether the timing discount is reachable. Halved rates only help if your work can wait. For an interactive feature it is irrelevant, and the off-peak column is the wrong one to budget against.
Where it fits
Reach for DeepSeek when cost dominates and the workload is schedulable: bulk classification, offline enrichment, evaluation runs, corpus processing. The combination of off-peak pricing and near-free cache hits makes those jobs far cheaper than the alternatives.
Reach elsewhere when latency is the constraint, when data residency rules it out, or when your evaluation set says the quality is not there for the task at hand.
Further reading
- DeepSeek, Pricing: the models, context, peak windows and the full rate table.
- Prompt caching: why the cache-hit rate changes prompt design.
- Cost and latency monitoring: recording cache fields separately.
- Choosing the right model: the step price pressure makes people skip.
Knowledge check
Question 1 of 3
Sign in to save your progress and pick up where you left off.