DeepSeek API pricing: what V4.1 Flash and V4 Pro cost, and when it is half price

DeepSeek API pricing has three moving parts: the model, whether the start of your prompt is already in DeepSeek's cache, and the time of day. On October 10, 2026, DeepSeek's own price list has two models, DeepSeek V4.1 Flash at $0.30 per million input tokens and $1.20 per million output tokens, and DeepSeek V4 Pro at $1.32 and $3.96. Those are peak-hour prices; outside peak hours everything costs half. This guide lists every price from DeepSeek's documentation, explains the cache and the peak window, works through what a real workload costs, and puts the result next to Claude, OpenAI and Gemini.

DeepSeek API prices per model

Table: DeepSeek API prices per million tokens (USD, checked October 10, 2026)

Both models have a 1 million token context window and up to 384,000 output tokens per request, both accept tool calls and JSON output, and both answer on an OpenAI-compatible endpoint (https://api.deepseek.com) and an Anthropic-compatible one (https://api.deepseek.com/anthropic). Only Flash reads images: DeepSeek lists vision as "not supported" on V4 Pro.

A few details the price table does not show:

• There is no batch discount. OpenAI, Anthropic and Google take 50% off work that can wait; DeepSeek's equivalent is the off-peak rate.
• The old Flash names still work. deepseek-v4-flash and deepseek-v4-flash-vision-exp are accepted, but those models are retired and the requests are served and billed as V4.1 Flash.
• V4 Pro was not switched off. DeepSeek's release post for V4.1 Flash on September 10, 2026 said all V4 Pro requests would be routed to Flash from September 14. Its change log adds that, "in response to user demand", it decided to keep serving V4 Pro after that date with billing unchanged. Pages that still say V4 Pro is gone were written from the first announcement.
• There is no free tier on the price list. Charges come out of a prepaid balance; if your account also has a granted balance (promotional credit), DeepSeek spends that first.

If you are new to the unit behind these prices, start with what a token is. Token counts are not comparable across vendors: each one splits text with its own tokenizer, so the same prompt has a different count on DeepSeek, Claude or GPT.

Peak and off-peak hours

DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Every other hour is off-peak, including whole weekends and whole Chinese public holidays, and off-peak prices are exactly half of peak. DeepSeek introduced the two rates with the V4 lineup in August 2026 and kept them when it cut prices for V4.1 Flash.

Converted to two common time zones while summer time lasts:

• London (BST, until October 25, 2026): 02:00 to 05:00 and 07:00 to 11:00 on weekdays. After the clocks change, London is on UTC and the hours are the UTC ones.
• New York (EDT, until November 1, 2026): 9 p.m. to midnight on Sunday to Thursday evenings, and 2 a.m. to 6 a.m. Monday to Friday.

In practice the peak window covers the Chinese working day and the European morning. A European or American team that runs nightly jobs, evaluations or bulk classification in its own afternoon or evening pays the off-peak rate without changing anything else.

How the cache changes the bill

Context caching is on by default for every account and needs no code changes. DeepSeek stores prompt prefixes on disk, and when a later request starts with a stored prefix, those tokens are billed at the cache-hit price. On V4.1 Flash a cache hit costs 2% of a cache miss ($0.006 against $0.30); on V4 Pro, about 3.3%.

The rule that decides a hit is stricter than on some other providers. DeepSeek's guide says each cached prefix is "an independent, complete unit" and a later request only hits the cache if it fully matches one of those units. Prefixes are saved at the end of each request's input, at the end of the model's output, at fixed intervals in very long texts, and whenever the system notices that several requests share the same beginning. In practice:

• Multi-turn chats hit the cache naturally, because each new turn starts with the whole previous conversation.
• Several questions about the same long document hit the cache once DeepSeek has seen the shared beginning, so put the document first and the question last.
• Anything that changes near the top breaks it, such as a timestamp or a user name inside the system prompt.

Every response reports prompt_cache_hit_tokens and prompt_cache_miss_tokens in its usage data, so you can see the real hit rate instead of guessing. Our guide to prompt caching explains how to order a prompt for hits on any provider.

Thinking is part of the output. Thinking mode is on by default on both models, at high effort; you can switch it off or set the effort to low or max. The chain of thought comes back as reasoning_content, and the API reference lists reasoning_tokens as a breakdown of the completion tokens, which are billed at the output price. For simple extraction or classification, switching thinking off is the quickest way to cut output.

What a real workload costs

Take the same support assistant we use in our other pricing guides: 100,000 questions a month, each sending 2,000 tokens of fixed instructions and reference text plus 1,000 new tokens, and getting back 400 tokens. With a steady flow the fixed 2,000 tokens are cache hits after the first requests.

On DeepSeek V4.1 Flash at peak rates the month costs:

• Cache-miss input: 100 million tokens × $0.30 per million = $30
• Cache-hit input: 200 million tokens × $0.006 per million = $1.20
• Output: 40 million tokens × $1.20 per million = $48
• Total: $79.20 a month, or $39.60 if all the traffic fell off-peak. Without any cache hits it would be $138.

Table: monthly cost of the sample workload (our calculation from each provider's prices, checked October 10, 2026)

Two things stand out. First, DeepSeek is no longer the cheapest API per token: Claude Haiku 5.5 ($0.10 input, $0.50 output) and GPT-6 Luna (the same list prices) come out below V4.1 Flash on this workload, even at off-peak rates. Anthropic warns that its newer tokenizer produces about 30% more tokens for the same text, which would put Haiku 5.5 nearer $41.60, still close to Flash off-peak. Second, the table compares prices, not quality. DeepSeek reports that V4.1 Flash beats its own V4 Pro on its benchmarks; we have not tested any of these models against each other, so run your own prompts before switching. Gemini 3.8 Flash's price doubles on January 1, 2027.

Cache hits matter most for agents. An agent that makes 20,000 calls a month, each re-reading 50,000 tokens of which 90% is cached and writing 1,000 tokens, costs $59.40 on V4.1 Flash at peak: $30 of uncached input, $5.40 for 900 million cached tokens and $24 of output. On V4 Pro the same month costs $250.80. The step-by-step math for every provider, with a calculator, is in our comparison of LLM API prices, and the single-provider guides cover Claude API pricing, OpenAI API pricing and Gemini API pricing.

Limits, data and how to keep the bill down

Concurrency, not tokens per minute. DeepSeek limits each account to 2,500 requests in flight at once on Flash and 500 on V4 Pro; over that, the API returns HTTP 429. Raising the limit is free on request. A user_id parameter separates your own customers for content safety, cache isolation and scheduling.

Where your data goes. DeepSeek's Open Platform terms (in force since April 29, 2026) say the personal data you give DeepSeek as a developer is handled under its privacy policy, and that for the data of your own users you are the controller. The privacy policy (last updated February 10, 2026) says DeepSeek collects, processes and stores personal data in the People's Republic of China and lists training and improving its models among its purposes; we did not find a separate statement on whether API inputs are used for training. It names Prighter Group as its privacy representative in the European Union and the United Kingdom. If you send customer data through the API, have whoever handles data protection in your company read both documents first.

Five ways to pay less:

1. Move flexible work off-peak. Reports, evaluations and bulk jobs scheduled outside the UTC windows cost half.
2. Keep the start of every prompt identical so the 2% cache-hit price applies, and check the hit rate in the usage data.
3. Turn thinking off where it adds nothing, and set a maximum output length.
4. Start on V4.1 Flash. Its input costs less than a quarter of V4 Pro's and its output less than a third, and DeepSeek itself ranks it higher on its benchmarks.
5. Compare the cost of a task, not the price per token, because each vendor counts tokens differently.

DeepSeek reserves the right to change its prices and recommends checking its pricing page regularly; we review this page every month.