What Does the DeepSeek API Cost in 2026?
DeepSeek-V4-Pro-0813 costs $0.66 per million input tokens on a cache miss and $1.98 per million output tokens off-peak, and exactly double that during peak hours. DeepSeek is the only major model vendor that charges a different price depending on what time of day you send the request.
Two multipliers decide your entire DeepSeek bill: the clock, and whether the input hit the cache. Every other pricing question about DeepSeek is downstream of those two.
| Model and window | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| V4-Pro-0813, off-peak | $0.022 | $0.66 | $1.98 |
| V4-Pro-0813, peak | $0.044 | $1.32 | $3.96 |
| V4.1-Flash, off-peak | $0.003 | $0.15 | DeepSeek has not published this |
| V4.1-Flash, peak | $0.006 | $0.30 | DeepSeek has not published this |
All figures are USD per million tokens, taken from DeepSeek's own API documentation in September 2026. DeepSeek prices aggressively and is also the one major provider with nothing to hand a startup, which is the gap AI Perks exists to fill.
How DeepSeek Peak and Off-Peak Pricing Works
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Every other hour, including all of Saturday and Sunday, is off-peak and costs half as much.
That is seven peak hours a day across five days, or 35 hours out of the 168 in a week. Derived from the published windows, 79% of the week is off-peak. The expensive rate is the narrow exception, not the default, which is the opposite of how most people read the announcement.
Convert the windows and the design becomes obvious. 01:00 to 04:00 UTC is 09:00 to 12:00 in Beijing, and 06:00 to 10:00 UTC is 14:00 to 18:00. DeepSeek's peak pricing is the Chinese working day, with the midday gap priced as off-peak. Nobody frames it that way, and it is the single fact that tells you whether the split affects you.
For a team serving the Americas it is close to a free discount. A 09:00 to 18:00 working day in US Eastern runs from roughly 13:00 to 22:00 UTC, which sits entirely inside off-peak. If your traffic follows US business hours, you are already on the lower rate and did nothing to earn it.
The teams that actually pay peak are the ones serving Asian daytime traffic, plus anyone running nightly batch jobs that happen to land in the 06:00 to 10:00 UTC window. Shifting that batch outside 06:00 to 10:00 UTC, into the 04:00 to 06:00 gap or to after 10:00, halves its bill with no code change.
Why a Cache Hit Beats the Clock
On V4-Pro, cached input costs $0.022 per million tokens against $0.66 for a cache miss. That is a 30x discount, versus 2x for timing. Off-peak halves your bill. Caching divides it by thirty.
Stack both and the spread on one model becomes extreme. A peak cache miss at $1.32 against an off-peak cache hit at $0.022 is a 60x difference on the same token, on the same model. On V4.1-Flash the spread is wider still: $0.30 peak cache miss against $0.003 off-peak cache hit is 100x. Both ratios are derived from the published rates.
Here is what that means on a realistic agent workload of 10M input and 2M output tokens a day, derived from the rates above:
| Daily cost line | Peak, no caching | Off-peak, 90% cache hit |
|---|---|---|
| 10M input tokens | $13.20 | $0.86 |
| 2M output tokens | $7.92 | $3.96 |
| Total per day | $21.12 | $4.82 |
| Total per 30 days | $633.60 | $144.54 |
The interesting part is not the 4.4x saving. It is the second column's shape: once input is cached, output is 82% of the bill. Output tokens get no cache discount at any tier, so past a certain caching quality the only remaining lever on DeepSeek spend is what time your requests arrive. Teams that optimise caching hard and then stop are leaving the second multiplier untouched.
If model spend is the line item you are trying to kill outright rather than shrink, credits from the providers that issue them are the faster route, and AI Perks tracks $7.7M in credits across 194 companies.
What Changed in DeepSeek Pricing This Year
The August 2026 change was not a price cut. It was a price split. Any DeepSeek guide published before 16 August 2026, including the earlier DeepSeek pricing page on this site, quotes one flat rate per model, and that model of DeepSeek's cost no longer describes the bill.
| Date | Release | Pricing effect |
|---|---|---|
| 24 April 2026 | V4-Pro and V4-Flash preview | 1M token context, up to 384K output, switchable thinking modes |
| 31 July 2026 | DeepSeek-V4-Flash-0731 | Flat rate: $0.14 cache-miss input, $0.28 output |
| 13 August 2026 | DeepSeek-V4-Pro-0813 general availability | Flagship reasoning and agent model |
| 16 August 2026 | Peak and off-peak split introduced | Time of day becomes a price multiplier |
| around 10 September 2026 | DeepSeek-V4.1-Flash | New input rates, older endpoint names retired |
Compare the Flash line before and after. The old flat rate was $0.14 per million cache-miss input tokens. The new off-peak rate is $0.15 and the peak rate is $0.30. Off-peak landed roughly where the flat rate already was, while peak is more than double it (derived). If your traffic concentrates in Asian business hours, your Flash bill went up in August even though no headline number got worse.
The April preview figures and the July flat rates are medium-confidence historical context here, not current pricing. Price anything you are actually shipping against the live docs.
The Endpoint Names DeepSeek Quietly Retired
DeepSeek-V4.1-Flash removed the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names. Any request still sending either string will fail.
This is the change that costs teams an afternoon, because it does not appear in a pricing table and it does not degrade gracefully. The call either resolves to a current model name or it errors.
Most tutorials, Stack Overflow answers and starter repos published before September 2026 still hardcode the retired names. If a DeepSeek integration that worked in August stopped working in September and nothing else changed, check the model identifier before you check anything else.
Does DeepSeek Offer Free API Credits or Startup Credits?
The DeepSeek chat app is free to use. The API has no stated free tier on DeepSeek's official pricing page, and DeepSeek runs no dedicated startup credit program at all.
Be careful with the second-hand version of this. Several sites state that new DeepSeek API accounts receive free starter credits. That claim is inconsistently reported and is not confirmed on DeepSeek's official pricing page, so treat it as unverified rather than as budget. The honest answer is that DeepSeek competes on list price, not on giveaways.
Third-party hosts are the other route in. Baseten lists DeepSeek V4.1 Flash input at $0.30 per million and carries a standing $30 new-workspace trial credit, which is the closest thing to free DeepSeek capacity with a real number attached. OpenRouter, Together and Fireworks also host DeepSeek models; their rates move often enough that you should read them yourself rather than trust a figure in an article. Routing through a host also flattens the peak and off-peak question, since the host sets its own rate card.
That leaves a genuine gap for founders. DeepSeek is among the cheapest ways to run inference in 2026 and the one major model provider that will not fund your first year. Almost every other provider on your shortlist does, which is why the practical stack is DeepSeek at list price for volume work plus funded credits everywhere else. AI Perks tracks which of those programs are open right now.
Frequently Asked Questions
What are DeepSeek's peak hours?
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything outside those windows, including all weekend, is off-peak and costs half as much. That works out to 35 peak hours in a 168 hour week, so roughly 79% of the time you are paying the lower rate.
How much does the DeepSeek API cost per million tokens?
DeepSeek-V4-Pro-0813 costs $0.66 per million input tokens on a cache miss and $1.98 per million output tokens off-peak. Peak rates are exactly double: $1.32 and $3.96. Cached input drops to $0.022 off-peak. Free credits for the providers that issue them are tracked at AI Perks.
Does DeepSeek have a free API tier?
The DeepSeek chat app is free to use. The API has no stated free tier on DeepSeek's official pricing page. Claims that new API accounts receive free starter credits are inconsistently reported and not confirmed officially, so do not build a budget around them. Verified credit programs from other providers are listed at AI Perks.
Does DeepSeek offer startup credits?
No. DeepSeek runs no dedicated startup credit program, which makes it unusual among major model providers. Baseten, which hosts DeepSeek models, carries a standing $30 new-workspace trial credit. For the providers that do run founder programs, AI Perks tracks $7.7M in credits across 194 companies at getaiperks.com.
Why did my DeepSeek API call stop working?
Most likely the model name. DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoint names, and any request still sending either string will fail. Older tutorials and code samples were written before the change and were never updated. Swap the model identifier for a current one and the call works again.
Is DeepSeek still the cheapest API in 2026?
On list price, off-peak cached input at $0.022 per million is hard for anyone to match. But the August split means a flat comparison against other vendors is no longer meaningful: your effective DeepSeek rate depends on when your traffic arrives and how well you cache. Compare against your own traffic pattern, not a table.
DeepSeek prices by the clock. Everyone else will just give you the credits.