How Much Does DeepSeek Off-Peak Pricing Save?
DeepSeek is the only major model vendor that charges by time of day. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and everything outside those windows costs exactly half.
That is seven peak hours per weekday, 35 hours out of the 168 in a week. About 79% of the calendar is already off-peak (derived from the published windows). Most teams pay the full rate on a fraction of their traffic and never notice they are doing it.
The discount is not a promotion or a committed-use tier. It applies automatically based on the clock when the request lands. AI Perks tracks provider pricing and credit programs across 194 companies, and this is the single most overlooked cost lever on any of them.
When Are DeepSeek Peak Hours in Your Time Zone?
The peak windows are fixed in UTC, so where they land in your working day depends entirely on where you sit. For North American teams, the entire business day is already off-peak.
Converted from the published UTC windows:
| Location | Peak window 1 | Peak window 2 |
|---|---|---|
| UTC | 01:00-04:00 | 06:00-10:00 |
| China (UTC+8) | 09:00-12:00 | 14:00-18:00 |
| India (UTC+5:30) | 06:30-09:30 | 11:30-15:30 |
| Central Europe (UTC+2) | 03:00-06:00 | 08:00-12:00 |
| UK (UTC+1) | 02:00-05:00 | 07:00-11:00 |
| US Eastern (UTC-4) | 21:00-00:00 | 02:00-06:00 |
| US Pacific (UTC-7) | 18:00-21:00 | 23:00-03:00 |
The China column explains the whole design. Both peak windows map onto the Chinese working day, morning and afternoon, with the lunch break sitting off-peak between them. Peak pricing is congestion pricing on DeepSeek's home market.
Two practical consequences:
If you are in the US, you are probably already off-peak by accident. Your 9-to-5 traffic sits outside both windows. What catches you is the overnight cron job, which is exactly when peak starts.
If you are in Europe, your morning is expensive and your afternoon is half price. The 08:00 to 12:00 local window covers the busiest part of a European working day.
The windows are fixed in UTC, so local times shift with daylight saving and a Sunday evening in the Americas can fall inside Monday's first peak window.
The Cache Multiplier That Stacks On Top
Off-peak halves your bill. A cache hit cuts input cost by roughly 30x. The two multiply, and together they account for almost the entire spread in what different teams pay DeepSeek for identical work.
Current published rates, in USD per million tokens:
| Model | Token type | Off-peak | Peak |
|---|---|---|---|
| DeepSeek-V4-Pro-0813 | Input, cache hit | $0.022 | $0.044 |
| DeepSeek-V4-Pro-0813 | Input, cache miss | $0.66 | $1.32 |
| DeepSeek-V4-Pro-0813 | Output | $1.98 | $3.96 |
| DeepSeek-V4.1-Flash | Input, cache hit | $0.003 | $0.006 |
| DeepSeek-V4.1-Flash | Input, cache miss | $0.15 | $0.30 |
A cached off-peak input token at $0.022 against a peak cache miss at $1.32 is a 60x spread on the same model. Nothing about the request changed except when it ran and whether the prefix was already resident.
Here is what that looks like on a real monthly workload of 50M input and 10M output tokens on V4-Pro (arithmetic derived from the table above):
| Configuration | Input | Output | Monthly total |
|---|---|---|---|
| Peak, no caching | $66.00 | $39.60 | $105.60 |
| Peak, 80% cache hits | $14.96 | $39.60 | $54.56 |
| Off-peak, no caching | $33.00 | $19.80 | $52.80 |
| Off-peak, 80% cache hits | $7.48 | $19.80 | $27.28 |
That is a 74% reduction for the same tokens and the same model. The middle two rows settle which lever to pull first: moving the clock alone ($52.80) edges out caching alone ($54.56), and it is one line of cron config rather than a prompt refactor.
Then note the detail almost nobody points out: at $27.28, output tokens are $19.80 of the bill, about 73%. Once you cache and shift off-peak, DeepSeek stops being an input-cost problem entirely and becomes an output-length problem. Trimming verbose responses is then worth more than any further prompt optimisation.
What Changed Since the Last DeepSeek Pricing Guide
DeepSeek-V4-Pro-0813 reached general availability on 13 August 2026, and time-of-day pricing arrived on 16 August 2026. Any DeepSeek cost guide written before those dates describes a pricing model that no longer exists.
The timeline that matters:
- 24 April 2026: V4-Pro and V4-Flash preview, with a 1M token context window, up to 384K output, and switchable thinking and non-thinking modes (confidence: medium).
- 31 July 2026: DeepSeek-V4-Flash-0731, the stable Flash release. Last known pricing before the change was $0.14 per million cache-miss input tokens and $0.28 per million output (confidence: medium, historical only).
- 13 August 2026: V4-Pro-0813 general availability.
- 16 August 2026: peak and off-peak pricing takes effect.
- Around 10 September 2026: DeepSeek-V4.1-Flash, live in the docs as DeepSeek-Flash.
Compare the old flat Flash input price of $0.14 against the new $0.15 off-peak and $0.30 peak. The August change was not a price cut, it was a price split. Run Flash off-peak and you pay roughly what you always did. Run it inside a peak window and you pay about double the old rate. Teams that never looked at the clock quietly took a price increase.
The Breaking Change in DeepSeek V4.1-Flash
DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoint names. Code written against them fails, and most tutorials still published online reference them.
This landed with no fanfare, which is the worst version of a breaking change: the tutorials, Stack Overflow answers and starter repos that rank well for DeepSeek queries were all written before it and have not been updated. If you are debugging an integration that worked last month, check your model string before you check anything else.
The lesson generalises: DeepSeek documents changes in the API docs rather than announcing them, so pin model identifiers explicitly and re-read the pricing page quarterly.
Does DeepSeek Offer Free Credits or a Startup Program?
The DeepSeek chat app is free to use. There is no stated API free tier on the official pricing page, and DeepSeek runs no dedicated startup credit program at all.
Be careful with what you read elsewhere here. Claims that new API accounts receive free credits are inconsistently reported and not confirmed on DeepSeek's official pricing page. Plenty of articles repeat a figure anyway. Treat any specific number you see as unverified until you have it from DeepSeek directly.
That absence is the real story for founders. Nearly every major provider now runs some form of credit program, and DeepSeek does not. Its answer to cost is low list pricing plus the off-peak and cache mechanics above, which is a genuinely different strategy, but it means there is no free runway to apply for.
The practical move is to treat DeepSeek as your cheap always-on tier and fund the expensive tiers with credits from the providers that do run programs. AI Perks tracks $7.7M in credits across 194 companies, which is where that runway actually comes from.
Find the programs that do give credits at getaiperks.com →
For third-party hosting, Baseten lists DeepSeek V4.1 Flash input at $0.30 per million and offers a standing $30 new-workspace trial credit (verify before budgeting). OpenRouter, Together and Fireworks also host DeepSeek models. Their prices move often enough that the routing decision should be made on latency, region and reliability, then re-priced at the moment you commit.
How to Restructure a Workload Around Off-Peak Hours
Step 1: Fund your premium tier first. Start at getaiperks.com and collect credits for the providers you route hard tasks to, so DeepSeek only carries the volume work.
Step 2: Move every batch job out of the peak windows. Backfills, embeddings runs, evaluation sweeps and nightly summarisation have no deadline. Shifting a cron job by two hours is a 50% discount for one line of config.
Step 3: Stabilise your prompt prefix. Cache hits depend on an identical leading prefix, so put the volatile parts of your prompt last. This is where the 30x lives.
Step 4: Measure your actual cache hit rate. Most teams assume caching is working and have never checked. The gap between an assumed 80% and an actual 30% is most of your bill.
Step 5: Attack output length. Once steps 2 and 3 are done, output tokens dominate. Shorter responses are now your biggest remaining lever.
Frequently Asked Questions
What are DeepSeek's off-peak hours?
Off-peak is everything outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which derives to roughly 79% of the week including all of Saturday and Sunday. Off-peak requests cost exactly half the peak rate on both input and output tokens, applied automatically with no configuration needed.
Does DeepSeek have a free API tier?
No. The DeepSeek chat app is free to use, but there is no stated API free tier on the official pricing page. Reports of free credits for new API accounts are inconsistent and unconfirmed. For actual free API runway, see the programs tracked at AI Perks.
How much cheaper is a DeepSeek cache hit?
On DeepSeek-V4-Pro-0813, a cache hit costs $0.022 per million input tokens off-peak versus $0.66 for a cache miss, which is 30x cheaper. Against a peak cache miss at $1.32, the spread reaches 60x on the same model for the same request.
Why did my DeepSeek API call stop working?
DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoint names. Older tutorials and code samples still reference them and now fail. Check your model identifier string first, since this change shipped quietly and most published guides have not been updated. Pin the current name explicitly in your config.
Does DeepSeek have a startup credit program?
No. DeepSeek runs no dedicated startup credit program, unlike most major model providers. Its cost strategy is low list pricing plus the off-peak and caching discounts instead. For providers that do offer startup credits, AI Perks tracks $7.7M across 194 companies.
Should I run DeepSeek through OpenRouter or directly?
Direct access is where the off-peak and cache pricing applies as published. Resellers set their own rates and may not pass the time-of-day discount through. Baseten, OpenRouter, Together and Fireworks all host DeepSeek models, so compare on latency and region, then verify the current price before committing.
DeepSeek will not give you credits. Plenty of others will, and they are all at getaiperks.com.