What a DeepSeek Cache Hit Actually Costs
On DeepSeek-V4-Pro, an input token served from cache costs $0.022 per million off-peak against $0.66 for a cache miss. That is a 30x discount for sending the same prefix twice.
Most pricing pages quote one input number and move on. DeepSeek has four, because two separate multipliers apply to every single call: whether the prefix hit cache, and what time of day you sent it.
| Model and window | Cache hit input | Cache miss input | Output |
|---|---|---|---|
| DeepSeek-V4-Pro-0813, off-peak | $0.022 | $0.66 | $1.98 |
| DeepSeek-V4-Pro-0813, peak | $0.044 | $1.32 | $3.96 |
| DeepSeek-V4.1-Flash, off-peak | $0.003 | $0.15 | not verified |
| DeepSeek-V4.1-Flash, peak | $0.006 | $0.30 | not verified |
All figures USD per million tokens, from the official API docs as of September 2026. V4.1-Flash output pricing is not verified here, so check the official page before budgeting on it.
The Flash ratio is even wider than the flagship's: $0.003 against $0.15 is a 50x gap (derived from the rates above). AI Perks tracks this kind of provider mechanic alongside the credit programs that offset it.
Peak and Off-Peak: The Only Vendor Charging by the Clock
DeepSeek peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Everything else, including all of Saturday and Sunday, bills at exactly half price.
No other major model vendor prices by time of day, which makes it the most overlooked lever in LLM cost work. No optimisation guide written for OpenAI or Anthropic has any reason to mention it.
Derived from those windows: peak is 7 hours a day across 5 weekdays, so 35 of the week's 168 hours. Roughly 79% of the calendar is already off-peak, so most teams get the discount by accident and could be getting more of it on purpose.
Where you sit decides how much of your workday is exposed:
| Team location | Peak windows, local time | Business-day exposure |
|---|---|---|
| US Eastern (UTC-4) | 21:00-00:00, 02:00-06:00 | None, the entire workday is off-peak |
| US Pacific (UTC-7) | 18:00-21:00, 23:00-03:00 | None |
| Central Europe (UTC+2) | 03:00-06:00, 08:00-12:00 | The whole morning |
| India (UTC+5:30) | 06:30-09:30, 11:30-15:30 | Most of the workday |
| China (UTC+8) | 09:00-12:00, 14:00-18:00 | Almost all of it |
Local times are converted from the published UTC windows at the offsets shown, and daylight saving moves them by an hour. The pattern is not an accident: the peak windows map onto the Chinese working day, so a US team's interactive traffic lands in the cheap half of the clock for free.
Stacking Both Discounts: A 60x Spread on One Model
A cached off-peak input token at $0.022 against a peak cache miss at $1.32 is a 60x spread on identical output from the same model.
Here is what that looks like on a realistic agent workload: a 200,000-token context resent 1,000 times a day, so 200 million input tokens daily on V4-Pro.
| Scenario | Rate per 1M | Daily input cost |
|---|---|---|
| Peak, every call a cache miss | $1.32 | $264.00 |
| Off-peak, every call a cache miss | $0.66 | $132.00 |
| Peak, every call a cache hit | $0.044 | $8.80 |
| Off-peak, every call a cache hit | $0.022 | $4.40 |
All four rows are derived by multiplying the published rates by 200. The gap between the top row and the bottom is about $260 a day, or roughly $95,000 a year, on a workload that does not change at all.
Two things follow that most cost write-ups miss:
A cached flagship beats an uncached small model. V4-Pro cached input at $0.022 off-peak is about 7x cheaper than V4.1-Flash uncached input at $0.15 (derived). Downgrading the model to save money is the wrong move if the real problem is a cache you never warmed.
Output is never cached. At $1.98 off-peak, an output token costs roughly 90x a cached input token (derived). Once caching works, you no longer have an input bill, you have an output bill, and every further saving comes from shorter responses or capped thinking budgets. Teams funding that output bill with provider credits can compare what is available at getaiperks.com.
What Changed Since Our Earlier DeepSeek Pricing Guide
Three things moved in 2026: V4-Pro-0813 reached general availability on 13 August, peak and off-peak pricing replaced flat rates in mid-August, and V4.1-Flash landed around 10 September.
Our earlier DeepSeek pricing page quotes flat per-token rates with no time-of-day component. Those predate the change and should be read as history.
The timeline, with confidence marked honestly:
- 24 April 2026 (medium confidence): V4-Pro and V4-Flash preview, with a 1M-token context window, up to 384K output tokens, and switchable thinking and non-thinking modes.
- 31 July 2026 (medium confidence): DeepSeek-V4-Flash-0731, the stable Flash release. Last known pricing before the change was $0.14 per million cache-miss input and $0.28 per million output, both flat. Historical only.
- 13 August 2026: DeepSeek-V4-Pro-0813 general availability, the current flagship reasoning and agent model.
- Around 10 September 2026: DeepSeek-V4.1-Flash, listed in the docs as "DeepSeek-Flash".
If you are working from a tutorial or a cost model written before August 2026, every input figure in it is wrong in both directions: too high for a cache hit, too low for a peak cache miss.
The Endpoint Rename That Breaks Old Code
DeepSeek-V4.1-Flash quietly retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names. Requests still passing those strings will fail.
This is the kind of change that surfaces as a production incident rather than a migration ticket, because the model name usually sits in a config file nobody has opened in months. Older code samples, SDK wrappers, and third-party tutorials still reference the retired names.
Worth doing before your next deploy:
- Grep the whole repo for both retired strings, including infrastructure config and notebooks
- Check any vendored SDK or agent framework that hardcodes a default DeepSeek model
- Confirm the vision path separately, since the experimental vision endpoint was renamed too
- Re-run your cost model afterwards, because the replacement sits on the new peak and off-peak schedule
DeepSeek Free Tier and Credits: What Is Actually True
The DeepSeek chat app is free to use. The API is not, and there is no stated free tier on the official pricing page.
This is where most articles get it wrong. Claims that new API accounts receive free credits circulate widely, are inconsistently reported, and are not confirmed on the official pricing page. If you need a starting balance, assume you are paying for it.
More important for founders: DeepSeek runs no dedicated startup credit program. There is no application, no stage-based tier, no partner track. Cheap per-token pricing is the entire offer, which is a real gap if you are assembling a credits stack.
That gap is exactly why the routing decision matters. Plenty of providers do run startup programs, and AI Perks tracks $7.7M in credits across 194 companies covering the ones that do. The practical pattern is to route high-volume, cache-friendly traffic to DeepSeek at $0.022 per million and spend granted credits on the pricier providers, where a free token is worth far more.
Buying DeepSeek From Someone Else
Third-party hosts do not replicate DeepSeek's peak and off-peak schedule, which means they can undercut the first-party API during peak and lose badly to it off-peak.
Baseten lists DeepSeek V4.1 Flash input at $0.30 per million (medium confidence, verify before you budget on it), and the output figure was not reliably confirmed. Baseten also carries a standing $30 new-workspace trial credit.
Notice what $0.30 equals: it is exactly DeepSeek's own peak cache-miss rate, and double the off-peak rate (derived). For a US or European team whose traffic mostly lands off-peak anyway, a flat third-party rate is not obviously a saving.
OpenRouter, Together and Fireworks also host DeepSeek models. We are not quoting their prices here because we have not verified them, and neither should anything else you read. The qualitative trade is consistent though:
- First-party API: the only place the full cache-hit and off-peak stack applies, and the cheapest path by a wide margin when both fire
- Third-party hosts: flat rates, different regions, unified billing, and usually no exposure to DeepSeek's cache-hit tier
- Aggregators: useful for failover, at a premium you should measure rather than assume
Price your own traffic against both, then compare the result with the credits available on other providers at getaiperks.com. A free token beats a cheap one.
Frequently Asked Questions
How much cheaper is a DeepSeek cache hit?
On DeepSeek-V4-Pro, a cache hit is $0.022 per million input tokens off-peak against $0.66 for a cache miss, a 30x difference. On V4.1-Flash the gap is wider still, $0.003 against $0.15, a 50x difference (derived). Rates are from the official API docs as of September 2026.
When are DeepSeek peak hours?
Peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays only. Everything else, including both weekend days, bills at half the peak rate. That works out to about 79% of the week being off-peak (derived), so US-based teams get the discount during their whole business day without changing anything.
Does DeepSeek have a free API tier?
No free API tier is stated on the official pricing page. The consumer chat app is free, but the API is paid from the first token. Reports of free credits for new accounts are inconsistent and unconfirmed. Providers that do grant credits are tracked at getaiperks.com.
Does DeepSeek offer startup credits?
DeepSeek runs no dedicated startup credit program. Low per-token pricing is the whole offer. For founders building a credits stack, that is a gap worth filling elsewhere: AI Perks covers $7.7M in credits across 194 companies from the providers that do run programs.
Why did my DeepSeek API call stop working?
The most likely cause is the model name. DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoint names, and older tutorials and config files still reference them. Grep your repository for both strings, including vendored SDKs and notebooks, then re-check your cost model against current rates.
Should I use DeepSeek V4-Pro or V4.1-Flash?
Run the arithmetic before assuming Flash is cheaper. V4-Pro cached input at $0.022 off-peak is roughly 7x cheaper than Flash uncached input at $0.15 (derived), so a well-cached flagship can beat an uncached small model on the input line entirely. Output volume, not model size, usually decides the bill.
Cache the prefix, ship off-peak, and let someone else's credits cover the rest.