What Does the DeepSeek API Cost in September 2026?
DeepSeek-V4-Pro-0813 costs $0.66 per million cache-miss input tokens and $1.98 per million output tokens off-peak, and exactly double that at peak. DeepSeek is the only major model vendor that prices by time of day, which is why any single "DeepSeek API price" you read is wrong at least part of the week.
Current published pricing, in USD per million tokens:
| Model | Input, cache hit | Input, cache miss | 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 confirmed |
| DeepSeek-V4.1-Flash, peak | $0.006 | $0.30 | Not confirmed |
Flash output pricing is marked unconfirmed rather than guessed, because DeepSeek's published page did not confirm it. Verify it yourself before you budget on it.
Those are DeepSeek's own rates. DeepSeek runs no startup credit program, which makes it the odd one out among frontier vendors: AI Perks tracks $7.7M in credits across 194 companies, and DeepSeek is not one of them.
Peak and Off-Peak: DeepSeek Bills You by the Clock
Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Everything else, the entire weekend included, is off-peak at half the price.
That is seven peak hours per weekday, 35 per week. Derived from those windows, about 79% of the week bills at the off-peak rate (133 of 168 hours).
Convert the windows to Beijing time and the logic becomes obvious. 01:00-04:00 UTC is 09:00-12:00 local, and 06:00-10:00 UTC is 14:00-18:00. DeepSeek's peak window is the Chinese working day with the lunch break cut out of the middle, which is derived arithmetic on the stated UTC windows rather than a claim DeepSeek makes.
For a team outside Asia that is close to free money. Derived from the stated windows and September 2026 offsets, here is how much of a 09:00 to 18:00 local workday lands inside the peak window:
| Team location | 09:00-18:00 local, in UTC | Working hours at peak price |
|---|---|---|
| US Pacific (UTC-7) | 16:00-01:00 | 0 of 9 |
| US Eastern (UTC-4) | 13:00-22:00 | 0 of 9 |
| UK (UTC+1) | 08:00-17:00 | 2 of 9 |
| Central Europe (UTC+2) | 07:00-16:00 | 3 of 9 |
| India (UTC+5:30) | 03:30-12:30 | 4.5 of 9 |
| China (UTC+8) | 01:00-10:00 | 7 of 9 |
A US team serving interactive business-hours traffic never pays DeepSeek's peak rate at all. A European team pays it for a third of the day and can move batch work out of the 06:00-10:00 UTC window to avoid it entirely, for zero engineering cost.
Cache Hits Are the Other Multiplier, and They Stack
A cache hit on DeepSeek-V4-Pro costs $0.022 per million input tokens against $0.66 for a miss, exactly 30x cheaper (derived). Stack that with off-peak and the spread between the best and worst input price on the same model reaches 60x.
Compare $0.022 per million for a cached off-peak token against $1.32 per million for a peak cache miss. Same model, same token, 60x apart.
Here is what the two multipliers do to a real bill. Take 50M input tokens and 10M output tokens per month on V4-Pro, with every figure derived from the prices above:
| Scenario | Input | Output | Monthly total |
|---|---|---|---|
| Peak, cache miss | $66.00 | $39.60 | $105.60 |
| Peak, cache hit | $2.20 | $39.60 | $41.80 |
| Off-peak, cache miss | $33.00 | $19.80 | $52.80 |
| Off-peak, cache hit | $1.10 | $19.80 | $20.90 |
Identical work, a 5x spread, and the only two variables are when you send the request and whether your prompt prefix is stable.
One more conclusion falls out of that table. Once your input is cached, output is roughly 95% of the bill ($19.80 of $20.90). Cache optimisation has a ceiling, and past it the only lever left is generating fewer output tokens.
What Changed in DeepSeek V4 Since the Last Pricing Guide
Any DeepSeek guide quoting one flat input price was written before 16 August 2026 and is now wrong in both directions: too high for off-peak, too low for peak.
| Date | Release | What changed |
|---|---|---|
| 24 April 2026 | V4-Pro and V4-Flash preview | 1M-token context, up to 384K output, switchable thinking and non-thinking modes (confidence: medium) |
| 31 July 2026 | DeepSeek-V4-Flash-0731 | Stable Flash line at a flat $0.14 cache-miss input, $0.28 output (historical, now superseded) |
| 13 August 2026 | DeepSeek-V4-Pro-0813 | General availability of the flagship reasoning and agent model |
| Around 10 Sept 2026 | DeepSeek-V4.1-Flash | Current Flash generation, live in the docs as "DeepSeek-Flash" |
The July to September Flash comparison is the cleanest illustration. A flat $0.14 per million cache-miss input tokens became $0.15 off-peak and $0.30 peak. Off-peak is roughly the old price, peak is more than double it (derived). Traffic landing in the peak window got 114% more expensive (derived: $0.30 against $0.14) while the headline stayed "cheap".
The Endpoint Rename That Breaks Old DeepSeek Code
DeepSeek-V4.1-Flash quietly retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names. Code and tutorials that still reference them will fail.
This is the most common reason a DeepSeek integration that worked in August stops working in September. It is not a gradual deprecation warning, it is a model identifier that no longer resolves.
Two rules follow from it:
Pin the model name in one place. If the string appears in six files, a rename is six edits and a production incident.
Treat any tutorial older than a month as a naming risk. This single release retired two identifiers at once, with no deprecation window to warn you. Copy the architecture from an old guide, then verify the model string against the current docs.
Is There a DeepSeek API Free Tier or Startup Credit?
The DeepSeek chat app is free. The API has no stated free tier on the official pricing page, and DeepSeek runs no dedicated startup credit program.
This is where most DeepSeek articles get it wrong. Claims that new API accounts receive free credits are inconsistently reported and are not confirmed on DeepSeek's official pricing page. Treat that as unverified rather than as a plan you can budget against.
The missing startup program is the bigger gap. Nearly every other frontier vendor, including Anthropic, OpenAI, Google and the major clouds, runs some form of credit program for early-stage companies. DeepSeek does not, which means the cheapest model on the market is also the one you cannot get for free.
That inverts the usual advice. Use free credits where they exist and pay DeepSeek's already-low rates where they do not. AI Perks tracks $7.7M in credits across 194 companies, covering the providers that do run a program.
The Levers That Actually Move a DeepSeek Bill
Price your own workload first. Multiply your expected monthly input and output tokens through the four scenarios above. If the spread is small, DeepSeek's timing mechanics do not matter to you and any provider will do.
Check where your traffic lands in UTC. US-based interactive traffic is already off-peak. European and Asian teams can halve that portion of the bill by shifting batch jobs out of 01:00-04:00 and 06:00-10:00 UTC.
Stabilise your prompt prefix. Cache hits are roughly 30x cheaper than misses on V4-Pro. A system prompt that changes on every call, for example because it interpolates a timestamp, throws that away.
Pin the model string and verify it against the current docs, because deepseek-v4-flash no longer exists.
Cover the rest of the stack with credits. DeepSeek will not give you any, but most of your other vendors will. Start at getaiperks.com.
On third-party hosting: OpenRouter, Together, Fireworks and Baseten all serve DeepSeek models. The tradeoff is structural rather than a matter of cents, because aggregators bill a flat rate, so routing through one usually means giving up the off-peak discount in exchange for reliability, a single invoice and a fallback when DeepSeek's own endpoint is slow. Baseten lists DeepSeek V4.1 Flash input at $0.30 per million and carries a standing $30 new-workspace trial credit, both unverified here, so confirm before budgeting. Note that $0.30 is exactly DeepSeek's own peak price, so a flat rate at that level only wins if your traffic would have been peak anyway.
Frequently Asked Questions
How much does the DeepSeek API cost?
DeepSeek-V4-Pro-0813 costs $0.66 per million cache-miss input tokens and $1.98 per million output tokens off-peak. Peak hours, 01:00-04:00 and 06:00-10:00 UTC on weekdays, cost exactly double. Cached input falls to $0.022 per million. Credits for the rest of your stack are tracked at getaiperks.com.
When are DeepSeek's peak hours?
Peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays. Everything else, including the entire weekend, is off-peak at half price. Derived from those windows, about 79% of the week bills at the cheaper rate, and the windows correspond to the Chinese working day either side of lunch.
Does the DeepSeek API have a free tier?
No API free tier is stated on DeepSeek's official pricing page. The DeepSeek chat app is free to use, but that is a separate product. Claims of free API credits for new accounts are inconsistently reported and unconfirmed. For credits that are confirmed, see getaiperks.com.
Does DeepSeek offer startup credits?
DeepSeek runs no dedicated startup credit program, which is unusual among frontier model vendors. The practical approach is to use free credits from the providers that do run one and pay DeepSeek's low rates for everything else. AI Perks tracks $7.7M across 194 companies.
Why did my DeepSeek API call stop working?
The likeliest cause is the model name. DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp identifiers, so code or tutorials written before September 2026 that reference them will fail with a model identifier that no longer resolves. Check the model string against the current docs before debugging anything else.
Is DeepSeek cheaper direct than through OpenRouter or Baseten?
Off-peak, usually yes, because aggregators bill a flat rate and cannot pass on a time-of-day discount. Baseten lists V4.1 Flash input at $0.30 per million, which matches DeepSeek's own peak price. Aggregators win on reliability and routing rather than on the off-peak rate.
The cheapest model on the market is the one nobody gives away. Budget accordingly.