DeepSeek API Pricing 2026: Cost per Million Tokens

DeepSeek API pricing changed three times in 2026: new V4 models in April, peak and off-peak prices in August, and V4.1 Flash in September. Most pages that rank for it still quote retired models or launch prices. This guide lists the current price per million tokens for each model, explains the peak hours and context caching that decide your real bill, and works through cost examples. Prices are from DeepSeek’s official pricing page, checked on 4 November 2026.

DeepSeek API prices per million tokens

DeepSeek currently offers two models through its API: deepseek-flash (DeepSeek V4.1 Flash) and deepseek-v4-pro (DeepSeek V4 Pro). Every price has an off-peak and a peak rate.

The old model names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but they are routed to V4.1 Flash and billed at Flash prices. deepseek-chat and deepseek-reasoner were retired on 24 July 2026.

DeepSeek peak and off-peak hours

Since 16 August 2026 DeepSeek charges full price only during Chinese working hours and half price the rest of the time. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday (9:00–12:00 and 14:00–18:00 Beijing time, excluding Chinese public holidays). Weekends are off-peak all day.

For developers in the Americas, normal working hours fall entirely in the off-peak window. In India and Europe, scheduling batch jobs for the afternoon or evening halves the bill.

How context caching lowers your DeepSeek bill

Caching is the biggest lever on DeepSeek’s prices: a cache hit on V4.1 Flash costs 50 times less than a miss. It works automatically, as described in the caching guide:

• It is on by default, and storing the cache is free. OpenAI and Anthropic charge for cache writes; DeepSeek does not.
• A hit needs the start of your prompt to match a stored prefix exactly. Keep the system prompt, tool definitions and long documents at the start and unchanged between calls.
• Caches are best-effort and are cleared within hours to days when unused.
• Each response reports prompt_cache_hit_tokens and prompt_cache_miss_tokens, so you can check your hit rate.

Chatbots with a fixed system prompt and coding agents that resend the same files every turn benefit most.

V4.1 Flash or V4 Pro: which DeepSeek model to use

V4.1 Flash is the default choice: it is three to seven times cheaper depending on the token type, accepts images, and allows five times more concurrent requests. V4 Pro is the much larger model (1.6 trillion parameters, 49 billion active, against 552 billion for V4.1 Flash) and is worth the price for harder reasoning, long multi-step agent tasks and difficult code.

Thinking mode and reasoning tokens

Thinking mode is on by default, with effort levels low, high (the default) and max. Reasoning tokens are reported inside completion_tokens and have no separate price, so in practice they are billed as output. When you use tools, earlier reasoning has to be sent back with the conversation, so it is billed again as input. Lower the effort level for simple tasks to save output tokens.

Search and Fetch tool fees

DeepSeek’s web Search and Fetch APIs cost $0.75 per 1,000 successful requests each; failed or empty requests are free. Since 31 October 2026 this also applies when the model calls them as server-side tools through Claude Code or Codex. Function calling with your own tools has no extra fee beyond the tokens.

DeepSeek API cost examples

These estimates use the official prices above and leave out tax.

1. A support chatbot with 1,000 conversations a day. Each conversation uses 2,000 input tokens (1,200 of them cached) and 400 output tokens. On V4.1 Flash that costs $0.73 a day at peak prices, or about $21.82 for 30 days. Fully off-peak it is $10.91 a month, and with traffic spread evenly across the week about $13.18. The same load on V4 Pro costs $40 to $81 a month.
2. A coding agent session. Fifty turns with about 70K tokens of context each add up to 3.4M cached tokens, 0.1M uncached tokens and 75K output tokens. That is $0.14 on V4.1 Flash at peak ($0.07 off-peak) and $0.58 on V4 Pro ($0.29 off-peak).
3. A batch translation job. 30M uncached input tokens and 3M output tokens cost $12.60 on V4.1 Flash at peak, or $6.30 off-peak.
4. What $10 buys on V4.1 Flash. About 66.7M uncached input tokens off-peak (33.3M at peak), or 16.7M output tokens off-peak. DeepSeek’s rule of thumb is that one English character is about 0.3 tokens.

DeepSeek API pricing vs OpenAI, Claude and Gemini

Sources: OpenAI, Anthropic and Google AI Studio pricing pages.

The honest reading: on list price alone, GPT-6 Luna and Claude Haiku 5.5 are now slightly cheaper than V4.1 Flash for uncached input and output. DeepSeek wins when most of your input is cached, when you run off-peak, and when you need a 1M token context at a low price. Against mid-range models such as Gemini 3.8 Flash, GPT-6.1 Sol or Claude Sonnet 5.5, both DeepSeek models are far cheaper. On the chatbot example above, GPT-6 Luna would cost about $8.76 a month, Gemini 3.8 Flash $65.70 and Claude Sonnet 5.5 $171.60.

How DeepSeek API billing works

• Prepaid balance. You top up first and usage is deducted from the balance; any granted balance is used before money you paid. When the balance runs out, requests fail with an HTTP 402 error.
• Payment methods. PayPal, bank cards, Alipay and WeChat Pay, plus bank transfer for companies. Crypto is not accepted.
• No expiry. Topped-up balance does not expire, and unused balance can be refunded, minus handling fees.
• No tiers. All accounts pay the same prices. Rate limits are counted as concurrent requests per account across all keys; going over returns a 429 error, and DeepSeek raises the limit for free on request.

What changed in DeepSeek API pricing in 2026

• 24 April: DeepSeek V4 preview, with launch prices far below today’s Flash output price.
• 24 July: deepseek-chat and deepseek-reasoner retired.
• 13 August: V4 Pro generally available.
• 16 August: peak and off-peak pricing introduced.
• 10 September: V4.1 Flash released as deepseek-flash at today’s prices; V4 Flash retired.
• 17 October: V4.1 Flash update (1017) at unchanged prices, and the Search and Fetch APIs.
• 31 October: Search and Fetch became billable when used through Claude Code and Codex.

If you see a page quoting $0.14 / $0.28 or $0.55 / $2.19 per million tokens, it describes the April 2026 preview or the old 2025 R1 model.

DeepSeek API privacy and data location

DeepSeek’s privacy policy states that it collects, processes and stores personal data in the People’s Republic of China, and its platform terms are governed by Chinese law. If your company cannot send data to China, DeepSeek models are also offered by Microsoft Azure AI Foundry and Google Cloud on their own infrastructure, under their own terms and prices.

Buying DeepSeek API credits with crypto

DeepSeek itself does not take crypto. AITopUp sells a DeepSeek API key as a prepaid token pack paid in USDT or other crypto, starting at $3.99 for 10M tokens with no expiry at the time of writing. The key comes from a third-party gateway provider, not from DeepSeek, and uses an OpenAI-compatible base URL, so it works in Cursor, VS Code, Codex, OpenCode or your own scripts. Usage on that key is counted with model multipliers: V4 Pro uses half a pack token per token and V4.1 Flash about a third, so a 10M pack covers roughly 20M tokens on V4 Pro or 33M on Flash. As with any gateway, your requests pass through the provider, so keep sensitive data out of it.

For other providers there is a Gemini API key on the same terms. To compare chat subscriptions instead of APIs, see our list of ChatGPT alternatives.