Is DeepSeek Free? App Versus API
The DeepSeek chat app is free to use. The DeepSeek API is not, and there is no stated free tier on its official pricing page.
That distinction is where almost every "is DeepSeek free" answer online goes wrong. People type the query after using the chat app, assume the same deal covers the developer API, then get a bill.
The API is cheap, not free. How cheap depends on two multipliers most write-ups skip entirely: what time of day you send the request, and whether your input hits the prompt cache. Between them they move the price of a single input token by 60x on the same model.
AI Perks tracks $7.7M in credits across 194 companies, which matters here for a reason covered below: DeepSeek is not one of them.
What Is Actually Free on DeepSeek Right Now
Free: the chat app. Not free: the API. Unconfirmed: new-account API credits.
The honest state of it, uncertainty marked rather than smoothed over:
| Surface | Free? | What you actually get |
|---|---|---|
| DeepSeek chat app | Yes | Free to use |
| DeepSeek API | No | No stated free tier on the official pricing page |
| New-account API credits | Unconfirmed | Widely repeated, not confirmed on DeepSeek official pricing page |
| Dedicated startup credit program | No | DeepSeek runs none |
The third row needs care. Plenty of articles assert a specific dollar figure of free API credit for new DeepSeek accounts. Those claims are inconsistently reported and not confirmed on the official pricing page, so treat any number as unverified until it shows up in your own console.
If you need a guaranteed credit balance before writing a line of code, you want a provider that publishes one. AI Perks lists the ones that do.
DeepSeek API Pricing: Why Time of Day Changes Your Bill
DeepSeek is the only major model vendor that charges by clock time. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Everything else is off-peak at half price.
Not a discount campaign, and not a batch queue you wait on. Same synchronous endpoint, half the price, decided by when the request lands.
Current published pricing, 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 |
Flash output pricing is left out on purpose: we could not confirm it against the official page, and a guessed number is worse than a gap.
Derived from those windows: peak covers 7 hours a weekday, 35 hours a week, roughly 21% of the clock. Meaning about 79% of all wall time is already off-peak - the half price is the default state, and peak is the exception you pay a penalty for wandering into.
The consequence: every queue-able workload belongs in that 79%. Nightly embedding runs, evals, backfills, document processing, synthetic data. None care about a few hours of delay, and all cost double if scheduled carelessly.
Cache Hits and the 60x Spread on One Model
A cache hit on DeepSeek-V4-Pro costs $0.022 per million input tokens off-peak. A cache miss at peak costs $1.32. That is a 60x spread on identical tokens hitting the identical model.
The two multipliers stack, and that stacking is the entire story of DeepSeek's cost profile:
- Cache hit vs cache miss on V4-Pro: 30x ($0.66 miss against $0.022 hit, off-peak, derived)
- Cache hit vs cache miss on V4.1-Flash: 50x ($0.15 against $0.003, off-peak, derived)
- Off-peak vs peak: exactly 2x on every line above
- Worst case against best case: 60x on the same input token
Flash has the wider cache multiplier of the two: the cheap model rewards caching proportionally harder than the flagship does.
What it means in money, on a job of 10M input plus 2M output tokens through V4-Pro (arithmetic derived from the table above):
| Scenario | Cost |
|---|---|
| Peak hours, no cache hits | $21.12 |
| Peak hours, 90% cache hit rate | $9.64 |
| Off-peak, no cache hits | $10.56 |
| Off-peak, 90% cache hit rate | $4.82 |
Same job, same model, same output. $21.12 against $4.82 is a 4.4x swing decided entirely by prompt structure and a cron expression. Teams migrating off pricier providers leave most of that on the table, then conclude DeepSeek was not as cheap as advertised.
Note the shape of the bill once caching works: at a 90% hit rate off-peak, output tokens are 82% of the cost ($3.96 of $4.82, derived). Past that point, optimising input further is wasted effort. Shorten outputs instead.
What Changed Since the V4 Preview
DeepSeek shipped three times in five months in 2026, and the pricing structure changed underneath the models. Anything written before 16 August is describing a system that no longer exists.
The sequence:
- 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, the stable Flash release. Last known pricing before the change was $0.14/M cache-miss input and $0.28/M output, flat, with no time-of-day component. (confidence: medium, historical only)
- 13 August 2026 - DeepSeek-V4-Pro-0813 reaches general availability as the flagship reasoning and agent model.
- 16 August 2026 - the peak and off-peak split arrives, and every price above this line stops describing what you pay.
- around 10 September 2026 - DeepSeek-V4.1-Flash, live in the docs as "DeepSeek-Flash".
That 16 August boundary is why older DeepSeek pricing content, including our own earlier guide, reads as a flat per-token table. There is no longer a single correct price per token to quote.
The Breaking Change: Retired Endpoint Names
DeepSeek-V4.1-Flash quietly retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names. Code written against them fails.
No loud deprecation notice, no long sunset window. Tutorials, blog posts and LLM-generated code samples from before September 2026 still reference the old identifiers.
Two habits catch it:
Pin the model string in one place. A single config constant makes a rename a one-line fix instead of a grep across the repo.
Alert on the error, do not retry it. A retired model name fails immediately and permanently, so a blind retry loop turns a five-minute fix into a silent outage.
Routing through an aggregator such as OpenRouter, Together or Fireworks moves that maintenance burden to them, at the cost of a margin on every token and one more hop of latency. A real trade, not an obvious win.
Where DeepSeek Credits Actually Come From
DeepSeek runs no dedicated startup credit program. If you want a credit balance rather than a low per-token price, it has to come from somewhere else.
That gap is the honest answer to "how do I get free DeepSeek credits." There is no DeepSeek program to point at. What exists instead:
| Route | What it gives you | Confidence |
|---|---|---|
| DeepSeek chat app | Free app usage, no API access | High |
| DeepSeek API direct | Low prices, no stated free tier | High |
| Third-party hosts (OpenRouter, Together, Fireworks) | DeepSeek models under a different billing relationship | Prices vary, verify before quoting |
| Baseten | Lists DeepSeek V4.1 Flash input at $0.30/1M, standing $30 new-workspace trial credit | Medium |
| Other model providers | Published startup credit programs, tracked on AI Perks | High |
Baseten's output figure is not quoted here because it was not reliably confirmed. A gap beats a guess.
The last row is where the real money is. Anthropic, OpenAI, Google, AWS, Azure and dozens of others do run startup credit programs, and a balance is worth far more than any per-token saving DeepSeek offers. AI Perks tracks $7.7M in credits across 194 companies, an average of roughly $39,700 per company (derived) - the providers that actually hand out balances.
The strongest 2026 setup is both: DeepSeek off-peak with a warm cache as the cheap default tier, funded credits from other providers for everything hard. That is how founders run a production AI stack near zero for a first year.
Frequently Asked Questions
Is DeepSeek free to use?
The DeepSeek chat app is free to use. The DeepSeek API is not free, and there is no stated free tier on its official pricing page. API pricing is genuinely low, starting at $0.022 per million cached input tokens off-peak on V4-Pro. Credit programs from other providers are tracked at AI Perks.
Does the DeepSeek API give free credits to new accounts?
Unconfirmed. Claims of a specific free credit amount for new DeepSeek API accounts are inconsistently reported across sites and are not confirmed on the official pricing page. Treat any figure you see as unverified until it appears in your own console, and budget as though there is none.
What are DeepSeek peak and off-peak hours?
Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Everything outside those windows is off-peak, priced at exactly half. Derived from those windows, roughly 79% of all wall time is off-peak, so half price is the default and peak is the exception worth scheduling around.
How much cheaper is a DeepSeek cache hit?
On DeepSeek-V4-Pro, a cache hit costs $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 50x. Stacked with off-peak pricing, the spread between best and worst case reaches 60x on the same input token.
Does DeepSeek have a startup credit program?
No. DeepSeek runs no dedicated startup credit program, which is why "free DeepSeek credits" searches rarely produce a real answer. Other model and infrastructure providers do run them. AI Perks tracks $7.7M in credits across 194 companies.
Will my old DeepSeek code still work?
Not necessarily. DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names, so older tutorials and generated code samples referencing them will fail. Pin the model string in one config constant, and alert on the error rather than retrying it, since retries never recover.
DeepSeek gives you a cheap price. Other providers give you a balance. Take both at getaiperks.com.