🐋 DeepSeek Token Calculator
What a prompt costs on DeepSeek. Two things make it unusual: an off-peak discount of 50% for most of the day, and a cache hit price of around 2% of a cache miss, which is the steepest cache discount of any provider here.
The off-peak discount and the cheapest cache in the business
DeepSeek prices differently from the Western providers in two ways that can both halve or better a bill. Rates vary by time of day, with most of the day charged at half price, and a cache hit costs around 2% of a cache miss — the steepest cache discount of any provider covered here.
Current prices
| Model | Input | Output | Cache read |
|---|---|---|---|
| DeepSeek Flash (V4.1) | $0.3 | $1.2 | $0.006 |
| DeepSeek V4 Pro | $1.32 | $3.96 | $0.044 |
How off-peak pricing works
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Everything else, including all weekend, is off-peak at half the rate. For scheduled or batch workloads that is a 50% saving available purely by choosing when to run, and the table above shows peak rates, so the off-peak figure is half what you see.
Cache hits cost almost nothing
A cache hit on DeepSeek Flash costs $0.006 per million against $0.30 for a miss — two percent. Combined with the off-peak discount it drops to $0.003. For an application with a large stable prompt, the marginal cost of input effectively disappears, which changes what architectures are affordable.
What the low price means in practice
DeepSeek Flash is roughly a twentieth of a Western flagship on input and still a fraction of mid-tier models. That makes it attractive for high-volume, low-stakes work: classification, bulk extraction, first-pass drafting. As with any provider, check where the data is processed and under which jurisdiction before sending anything sensitive — that is a procurement question, not a pricing one.
Naming and retirement
The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but now resolve to the current Flash model and bill at its price. If you pinned an old name expecting old behaviour, you are getting the new model either way, which matters more for output consistency than for cost.
Prices come from DeepSeek API docs and were checked on 2026-10-26. AI pricing changes frequently — confirm against the provider before committing to a budget.
Once you have this figure, the ai token calculator is where it usually goes next. the grok token cost is the related one people reach for afterwards.
Frequently Asked Questions (FAQs)
How much does the DeepSeek API cost?
At peak rates on 2026-10-26, DeepSeek Flash was $0.30 per million input on a cache miss and $1.20 output; V4 Pro was $1.32 and $3.96. Off-peak, all of those halve.
When are DeepSeek's off-peak hours?
Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Every other hour, including the whole weekend, is off-peak at half price.
How cheap are DeepSeek cache hits?
About 2% of a cache miss: $0.006 per million against $0.30 on Flash at peak, falling to $0.003 off-peak. It is the steepest cache discount among the providers covered here.
Is DeepSeek cheaper than GPT or Claude?
Substantially, on list price. Flash is roughly a twentieth of a Western flagship on input. Whether that is the right trade depends on the task and on your data governance requirements, not on the price alone.
What happened to deepseek-v4-flash?
The name still works but now routes to the current V4.1 Flash model and bills at its rate. The older vision-experimental name does the same. Pinning the legacy name no longer pins the old behaviour.
Can I use DeepSeek for sensitive data?
That is a procurement and compliance question rather than a pricing one. Check where requests are processed, under which jurisdiction, and what the retention terms say, exactly as you would for any provider.
Are these prices current?
They reflect published figures checked on 2026-10-26. DeepSeek's own documentation site did not expose a machine-readable pricing table at the time, so confirm against the provider before relying on them.