Practical guide. Prices checked on September 6, 2026. All amounts are in US dollars.
When you use an AI service billed by the token, the amount of content it processes affects your costs. On Ollama Cloud, timing matters too: DeepSeek-V4-Flash and DeepSeek-V4-Pro have token rates that are 50% lower outside peak hours. If a task can wait, moving it to a cheaper window can make your usage credits go further. Ollama pricing.
This guide explains when to run those tasks, how to estimate the savings, and how to start using DeepSeek through Ollama. For subscription allowances and billing, see our guide to Ollama Cloud pricing and credits.
DeepSeek’s peak and off-peak token prices
The following rates apply to one million tokens, not one million requests.
| Model | Period | Input | Cached input | Output |
|---|---|---|---|---|
| DeepSeek-V4-Flash | Off-peak | $0.22 | $0.007 | $0.66 |
| DeepSeek-V4-Flash | Peak | $0.44 | $0.014 | $1.32 |
| DeepSeek-V4-Pro | Off-peak | $0.66 | $0.022 | $1.98 |
| DeepSeek-V4-Pro | Peak | $1.32 | $0.044 | $3.96 |
Source: Ollama’s official pricing table.
Input is the content supplied to the model. Output is the content it generates and bills for. Cached input refers to the portion eligible for the service’s discounted reuse rate. For an initial budget, assume uncached input unless you have evidence that your workload benefits from caching.
When does off-peak pricing apply?
Ollama lists peak hours as 12:00–18:00 UTC, Monday through Friday. Off-peak rates apply outside that window and throughout the weekend in UTC. Ollama’s peak pricing conditions.
For readers working in Central European time, the schedule converts as follows:
| Time reference | Weekday peak window | Weekday off-peak hours |
|---|---|---|
| UTC | 12:00–18:00 | Before 12:00 and after 18:00 |
| Paris, summer time — CEST, UTC+2 | 14:00–20:00 | Before 14:00 and after 20:00 |
| Paris, winter time — CET, UTC+1 | 13:00–19:00 | Before 13:00 and after 19:00 |
Time conversions calculated by DemystIA from Ollama’s published UTC window.
For recurring jobs, using UTC in your scheduler avoids seasonal changes relative to Ollama’s pricing window. Allow enough time for the whole workload to finish within your chosen period. The sources reviewed do not specify how a request that crosses a pricing boundary is billed.
What could you save on a batch of work?
Consider a hypothetical batch consuming two million uncached input tokens and 500,000 billable output tokens.
| Model | Entire batch at peak rates | Entire batch at off-peak rates | Saving |
|---|---|---|---|
| Flash | $1.54 | $0.77 | $0.77 |
| Pro | $4.62 | $2.31 | $2.31 |
DemystIA’s off-peak calculation for Flash is 2 × $0.22 + 0.5 × $0.66 = $0.77. The other totals use the same method. These are illustrative calculations at identical token volumes, not measurements of model productivity.
Across 100 identical batches, the theoretical saving would be $77 for Flash or $231 for Pro in metered usage. That does not reduce the subscription’s fixed price: it reduces the credits consumed by that work.
The comparison also does not establish which model is the better choice for your task. If one requires more retries or produces substantially more output, its actual cost will differ from this fixed-volume example.
How to run DeepSeek through Ollama
Ollama Cloud runs models remotely, allowing you to use models that exceed your computer’s available GPU resources. You need an Ollama account and an internet connection. Ollama Cloud documentation.
With Ollama installed, open a terminal and sign in:
ollama signinTo start Flash, run:
ollama run deepseek-v4-flash:cloudTo start Pro, run:
ollama run deepseek-v4-pro:cloudThese model identifiers are listed on the official DeepSeek-V4-Flash and DeepSeek-V4-Pro pages. The :cloud suffix selects remote execution; these commands do not run the full model on your PC.
Free accounts can purchase usage credits to unlock all models. Access through a free account does not mean unlimited free inference, so check your available balance and model access before starting a large job. Ollama’s usage conditions.
Which tasks are worth moving?
Good candidates are tasks whose results are not needed immediately: summarizing a document collection, preparing translation drafts, analyzing files, or running a set of evaluations.
Start with a small sample before scheduling the full workload. Record output quality, failures, and total consumption across all attempts. A low-cost generation that needs several corrections may offer less value than its initial price suggests.
For interactive troubleshooting or an urgent question, compare the expected saving with the cost of waiting. On small workloads, postponing useful work may save only a few cents.
What does Zero Data Retention mean here?
Ollama says it does not log prompts or responses or use them for training, and requires zero-retention policies from its hosting partners. Its general FAQ describes infrastructure primarily in the United States, with additional capacity in Europe and Singapore. Ollama’s privacy FAQ.
The off-peak announcement describes DeepSeek hosting in the US and Europe. This does not establish a guarantee that requests stay within Europe. Provider-side retention policies also do not describe any conversation history your own client application may keep.
Make scheduling part of your usage budget
For flexible workloads, off-peak pricing offers a straightforward opportunity: the same billable token volume costs half as much during the cheaper window. The practical next step is to measure one representative task, decide whether it can be delayed, and schedule it with enough margin to finish before peak pricing returns.
Keep the official pricing page as your reference when planning future runs, since rates and conditions can change.