DeepSeek API: how to get a key, what it costs per 1M tokens, and the truth about "free"

DeepSeek API pricing at a glance

DeepSeek's own pricing page, read 29 Oct 2026, in USD per 1M tokens. Off-peak is every hour outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so the whole US working day.

Billing is "number of tokens × price" from a prepaid balance, with no monthly fee and no free tier we could find. The full price section below covers cache hits, concurrency limits and the Novita AI comparison.

How to get a DeepSeek API key

DeepSeek's quick-start page lists one value you need for the API key field: "apply for an API key". It links to the platform. Here are the steps, using the page names from DeepSeek's own docs and FAQ:

1. Create an account at platform.deepseek.com. DeepSeek's FAQ says it may reject some email domains and recommends "a major international email provider such as Gmail, Outlook, Hotmail, or Yahoo". There is also a Google login, which the FAQ says does not work from Mainland China IP addresses.
2. Top up. On the Top Up page, the FAQ lists "PayPal, bank card, Alipay, or WeChat Pay". It also says "Top-up balances do not expire" and that unused balances can be refunded.
3. Create the key on the API Keys page, then put it in an environment variable called DEEPSEEK_API_KEY, which is the name DeepSeek's own code samples use.
4. Point your code at DeepSeek. The API works with the OpenAI and Anthropic formats. The base URL is https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic (Anthropic format), and the model names are deepseek-flash and deepseek-v4-pro.

If a request fails, two error codes cover most new accounts. 401 means the key is wrong. 402 "Insufficient Balance" means you have run out of balance, and DeepSeek's fix is to "go to the Top up page to add funds". A brand-new account with no top-up is the usual cause.

Is the DeepSeek API free?

No. On 29 Oct 2026 we read DeepSeek's pricing page, quick start, rate-limit page, error codes and FAQ, and none of them offers a free tier, a trial or sign-up credits. Every request is charged as "number of tokens × price".

The pricing page does mention a "granted balance": fees come out of "your topped-up balance or granted balance, with a preference for using the granted balance first". None of the pages we read says who gets a granted balance, how much it is, or whether a new account gets one. So don't plan around it. If your account shows one, it gets used before your own money.

Searches for "free DeepSeek API" usually turn up one of three things:

• The open weights. These really are free to download. Running them is not free, because you need the hardware (more on that below).
• Third-party "free API" endpoints and proxies. Your prompts go to whoever runs the endpoint, not to DeepSeek. Treat anything you send as readable by that operator.
• Sites offering "free DeepSeek API keys". DeepSeek only issues keys from inside your own platform account. A key someone else gives you belongs to someone else's account. It is either shared, which DeepSeek's FAQ says never to do, or leaked. The owner can revoke it at any time, and everything you send runs through an account you don't control. Don't use these keys, and never pay for one.

DeepSeek API price per 1M tokens

The table below is DeepSeek's own pricing page, read on 29 Oct 2026. Both models have a 1M-token context and a maximum output of 384K tokens.

Peak and off-peak. DeepSeek's wording is: "Off-peak rates are half of the peak rates. Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays." Weekends are off-peak all day. Peak time ends at 10:00 UTC, which is 6am in New York (5am once US clocks change on 1 Nov). So if you work US hours, most of your calls are billed at the off-peak rate without you doing anything.

Cache hits are input tokens DeepSeek has already seen, such as a long system prompt you send with every request. They cost about 2% of the cache-miss price. That makes a fixed prompt prefix the cheapest thing to send repeatedly.

Rate limits. Going over the concurrency limit returns an HTTP 429. DeepSeek says you can ask for more capacity and "There is no additional cost for capacity expansion."

Where your data goes. DeepSeek's privacy policy (last updated 10 Feb 2026) says: "we directly collect, process and store your Personal Data in People's Republic of China." For some teams that settles the question. The next section covers the alternative.

Can't or don't want to use DeepSeek's platform? Novita AI runs the same models

DeepSeek publishes the weights, so other companies can serve the same models through their own APIs. Novita AI is one of them. You sign up with Google, GitHub or email, create a key under Key Management, and call an OpenAI-compatible endpoint at https://api.novita.ai/openai/v1. Its quickstart sample uses deepseek/deepseek-v4.1-flash.

Here are Novita's prices for the same model versions, read on 29 Oct 2026 from its pricing page and public model list:

DeepSeek's own API is never more expensive. Novita charges DeepSeek's peak price at all hours. Off-peak, DeepSeek charges half. If all you want is the cheapest DeepSeek tokens and DeepSeek's sign-up and data terms work for you, use DeepSeek directly.

Novita makes sense for other reasons, each of which comes from its own pages:

• Not a DeepSeek account. Novita's privacy policy (effective 13 May 2026) is written for US privacy law, with a CCPA section. It says data may be transferred to "countries outside your country or region, including the United States". It does not say where inference runs. If data location is a hard requirement for you, ask Novita before you send anything sensitive.
• Older DeepSeek models. DeepSeek's price list now has two models. Novita still lists DeepSeek R1 0528 (160K context), so code that depends on R1 keeps working.
• One key for many open models. Novita's model list has 99 entries on 29 Oct 2026, including Llama, Qwen, Gemma, GLM and gpt-oss. You can switch models without opening a new account.
• Batch jobs. Novita's pricing page says "Batch inference is available at an introductory 50% discount on input and output tokens for supported models." Check which models qualify before you count on it.

Disclosure: Novita AI links on this page are referral links. If you sign up through one, we earn a commission on what you spend there. DeepSeek pays us nothing, and we recommend its own API wherever it is the cheaper choice.

The free route: run DeepSeek yourself

The model weights are the only part of DeepSeek that is genuinely free. DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro are both on Hugging Face under the MIT licence. But "free to download" is not "free to run". The full V4 models are datacenter-sized: our V4 Flash hardware breakdown puts a self-hosted setup at about $10k of GPUs.

The models that fit on a normal PC are the R1 distills, which are smaller models trained to reason like R1:

• How to run DeepSeek on your own computer, free walks through setup.
• Which DeepSeek R1 model fits your GPU matches each size to your VRAM.
• What it costs to run DeepSeek R1 Distill 32B compares buying, renting and electricity using live prices.

Running locally there is no key, no per-token bill, and your prompts stay on your machine. A distill is not the full model, though. For the quality of the full model, the cheapest option is still DeepSeek's off-peak price.

Keep your DeepSeek API key safe

A DeepSeek key spends your balance. Anyone who has it can drain your top-up.

• Server side only. DeepSeek's FAQ: "never share them with others or expose them in browsers or client-side code." If your app is a Lovable, Bolt or Next.js front end, the call has to go through a server function. Your API keys are in the browser shows how.
• Keep it out of git. Put it in an environment variable and keep .env out of your repo. Watch out for VITE_ and NEXT_PUBLIC_ prefixes, because they ship the value to the browser. The one file you must never commit explains why.
• If it leaks, revoke it. DeepSeek's FAQ steps: open the API Keys page, click the trash-bin icon next to the key, then confirm with "Revoke". Then create a new key and swap it into your app. The Usage page can export usage per key, which shows you what the leaked key was used for.

Common questions