DeepSeek API Pricing 2026: deepseek-flash and V4 Pro per Million Tokens, Peak vs Off-Peak, and Cost per Conversation

Published 27 October 2026. Every price below was read from DeepSeek's Models & Pricing page and DeepSeek's linked API documentation on 27 October 2026, and every price is as of that date.

DeepSeek API pricing has two models and two clocks, as of 27 October 2026. deepseek-flash (DeepSeek-V4.1-Flash-1017) costs $0.15 per million input tokens on a cache miss, $0.003 on a cache hit and $0.60 per million output tokens off-peak, and exactly double that ($0.30, $0.006 and $1.20) during peak hours. deepseek-v4-pro (DeepSeek-V4-Pro-0813) costs $0.66, $0.022 and $1.98 off-peak, and $1.32, $0.044 and $3.96 at peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, which is roughly 21% of the hours in a week, and every other hour, including weekends and Chinese public holidays, bills at the off-peak rate.

Two things on the incumbent pages are out of date. DeepSeek announced on 10 September that every deepseek-v4-pro request would be routed to V4.1 Flash from 14 September, and several pricing guides still say so; DeepSeek's own change log reversed that the same week, and V4 Pro is still served and billed at its own rates today. And DeepSeek is no longer the cheapest list price per token: Claude Haiku 5.5 lists at $0.10 input and $0.50 output, below deepseek-flash even off-peak. Where DeepSeek still wins is the cache hit, at $0.003 to $0.006 per million, and in the worked example further down a 40-turn conversation on deepseek-flash costs $0.026 at peak and $0.013 off-peak with caching, against $0.240 on Claude Sonnet 5.5 and $0.111 on Gemini 3.8 Flash.

DeepSeek API price per million tokens, every current model

US dollars per million tokens from DeepSeek's pricing page, as of 27 October 2026. DeepSeek bills "the total number of input and output tokens", so reasoning tokens bill as output. There is no separate cache write or cache storage charge.

Both models share one base URL (https://api.deepseek.com for the OpenAI format and https://api.deepseek.com/anthropic for the Anthropic format), a 1M-token context and a maximum output of 384K tokens, and both support JSON output, tool calls, the OpenAI Responses API, the Anthropic API, chat prefix completion (beta) and FIM completion (beta, non-thinking mode only). Image input is supported on deepseek-flash and not on V4 Pro. Each account can run 2,500 concurrent requests on deepseek-flash and 500 on V4 Pro, according to DeepSeek's rate limit page.

V4 Pro costs 4.4 times as much as deepseek-flash on uncached input, 3.3 times on output and 7.3 times on cache hits, at either hour. A cache hit costs 1/50 of a miss on deepseek-flash and 1/30 on V4 Pro.

If your account settles in yuan, the Chinese pricing page lists deepseek-flash at 0.02 yuan (cache hit), 1 yuan (cache miss) and 4 yuan (output) per million tokens off-peak, and 0.04, 2 and 8 yuan at peak; V4 Pro is 0.15, 4.5 and 13.5 yuan off-peak and 0.30, 9.0 and 27.0 yuan at peak.

DeepSeek peak and off-peak hours in your time zone

DeepSeek defines peak as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, "excluding Chinese public holidays", which is 09:00 to 12:00 and 14:00 to 18:00 Beijing time. Off-peak rates are half of peak on every meter, input and output alike.

Until US clocks change on 1 November, the US rows run one hour later (21:00 to 00:00 and 02:00 to 06:00 Eastern). The UTC boundaries never move.

India gets the worst of this schedule. A 09:30 to 18:00 IST working day overlaps peak for four of its eight and a half hours, and the 06:30 block catches early-morning traffic, so an Indian product with daytime users should budget closer to the peak column. A US product sees the opposite: a 09:00 to 17:00 working day in Eastern or Pacific time never touches a peak block, and its peak exposure is the evening and overnight hours from Sunday night to Friday morning, US time, which are weekday business hours in Beijing. A product whose traffic is spread evenly across the week pays the peak rate on roughly 21% of its tokens, which puts the blended deepseek-flash rate at about $0.18 for uncached input and $0.72 for output.

For anything that can wait, the move is to schedule it outside the two blocks. DeepSeek's documentation lists no Batch API, so the off-peak window is the discount for deferrable work: an evaluation run or a nightly back-fill started at 10:00 UTC instead of 08:00 UTC costs half as much with no code change.

Is DeepSeek V4 Pro still available?

DeepSeek V4 Pro is still on the API. When DeepSeek released V4.1 Flash on 10 September 2026, its announcement said that from 04:00 UTC on 14 September "all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates", and pricing guides published that week repeated it. DeepSeek's change log entry for the same release now says the opposite: "In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged." As of 27 October the pricing page still lists deepseek-v4-pro as DeepSeek-V4-Pro-0813 at its own rates, and a request to it is billed at V4 Pro prices, not Flash prices.

Which one to pay for is a question for your own evals. DeepSeek's V4.1 Flash notice says that "tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime", and DeepSeek then kept V4 Pro "in response to user demand". Neither is evidence about your prompts, so run yours on both models before paying 4.4 times as much for input.

Old model names and what they bill

Sources: pricing page footnote and V4 release notice.

What about DeepSeek V3.2 pricing?

DeepSeek V3.2 is no longer on DeepSeek's own API. It was served through deepseek-chat and deepseek-reasoner from 1 December 2025, and those two names were retired on 24 July 2026, so a search for "DeepSeek V3.2 pricing per million tokens" today has no first-party answer. The open weights are still hosted elsewhere: on 27 October 2026, OpenRouter's V3.2 endpoint list shows providers such as Baidu at $0.28 input and $0.42 output and GMICloud at $0.209 and $0.31, with a 163,840-token context, and the Amazon Bedrock pricing page lists DeepSeek V3.2 among its DeepSeek models. A new project starting today gets a 1M-token context and a lower uncached rate on deepseek-flash off-peak.

How DeepSeek context caching works, and what it bills

DeepSeek's context caching is on for every account by default and needs no code. Each request builds a disk cache, and when a later request starts with a prefix the cache already holds, those tokens bill at the cache-hit rate. You never pay to write a cache and never pay to store one, which is the main structural difference from Claude, where a five-minute cache write costs 1.25 times the input price, and from Gemini's explicit caches, which charge storage by the hour.

Matching is stricter than on most providers. DeepSeek stores "cache prefix units", and a later request hits only if it fully matches one of them. Units are created at three points: at the end of each request's user input and at the end of the model's output, whenever the system detects a common prefix across several requests, and at fixed token intervals inside long inputs and outputs so that a very long prefix is not left uncacheable. Multi-turn chat and shared documents behave differently under these rules:

• Multi-turn chat hits from the second turn: a second request of A + B + C fully matches the unit A + B that the first request created.
• Shared documents hit from the third request: two requests that share a long document but ask different questions (A + B, then A + C) both miss; the system then persists A as its own unit, and the third request (A + D) hits on it.

Caches are best-effort and short-lived: DeepSeek does not guarantee a 100% hit rate, builds a cache in seconds, and clears an unused one "usually within a few hours to a few days".

The pricing page lists no minimum cacheable length. Check what you were billed for in the usage object of every response: prompt_cache_hit_tokens and prompt_cache_miss_tokens add up to prompt_tokens, according to the Chat Completions reference.

Two settings change how the cache behaves. Passing a user_id isolates the KV cache per end user on your side, which DeepSeek describes as a privacy control on its rate limit page, so expect a shared system prompt to be cached per user_id rather than once for the whole account. And on the Anthropic-format endpoint, cache_control is ignored, according to DeepSeek's Anthropic API guide: code written for Claude's explicit breakpoints still runs, and caching happens automatically instead.

Reasoning tokens: on by default, billed as output

Thinking mode is enabled by default on both models, at "high" effort, according to DeepSeek's thinking mode guide. Reasoning tokens are reported as completion_tokens_details.reasoning_tokens inside completion_tokens, so they bill at the output rate. A request whose visible answer is 300 tokens and whose reasoning is 2,000 tokens bills 2,300 output tokens, which is $0.0028 on deepseek-flash at peak instead of $0.00036.

You set effort with reasoning_effort (low, high or max) in the OpenAI format or output_config.effort in the Anthropic format, and you turn thinking off with {"thinking": {"type": "disabled"}}. Other effort names are mapped: minimal runs as low, medium and xhigh run as high, and ultra runs as max.

The less obvious cost sits in agent loops. In a plain chat, DeepSeek ignores the reasoning_content of earlier turns, so yesterday's reasoning is not re-billed as input. Once a request carries the tools parameter, the guide says "the reasoning_content of all previous turns should be passed back to the API and will be concatenated into the context", and omitting it returns a 400 error. In a tool-using agent, every reasoning token is therefore billed once as output and again as input on every later step, at the cache-hit rate when the prefix matches and at the full rate when it does not.

What a multi-turn conversation costs on DeepSeek, Claude and Gemini

Take a chat assistant with a 2,000-token system prompt, where each user message is 150 tokens and each reply is 350 tokens; these are the same token assumptions as our Gemini API pricing and Claude API pricing guides, so the numbers line up across all three. Turn 1 sends 2,150 input tokens. Turn 40 sends 21,650, because the API is stateless and every request carries the 39 exchanges before it, and the whole 40-turn conversation consumes 476,000 input tokens to produce 14,000 output tokens. A 100-turn conversation consumes 2,690,000 input tokens. Reasoning tokens are left out, so thinking mode would add to every output figure below.

Rates for the comparison come from Anthropic's pricing page (Haiku 5.5 at $0.10 input for prompts up to 100,000 tokens, $0.50 output, $0.125 five-minute cache write and $0.01 cache read; Sonnet 5.5 at $2, $10, $2.50 and $0.10) and Google's Gemini API pricing page (3.8 Flash at $0.75, $3.75 and $0.075 cached until 31 December, doubling on 1 January; Argon at $2, $10 and $0.10 until 30 November), all read on 27 October 2026.

"Cache warm" is modelled the same way for DeepSeek as for Gemini's implicit cache on our Gemini page: from turn 2, the previous request's prompt is read at the cache-hit rate and the new 500 tokens (last reply plus new message) are billed as a miss. That is the conservative case for DeepSeek. In non-thinking mode the unit persisted at the end of the model's output can also cover the previous reply, leaving only the 150-token user message as a miss, which brings the 40-turn deepseek-flash figure down to $0.022 at peak and $0.011 off-peak. For Claude, each turn writes its new tokens at 1.25 times the input price and reads everything before them, from turn 1, since Claude caches from 512 tokens according to Anthropic's caching guide. Gemini's implicit cache starts once the previous request passes 4,096 tokens, and Argon re-processes 2,050 tokens on every short turn; both rules are worked through on our Gemini page.

Four things come out of the table. Without caching, Claude Haiku 5.5 is the cheapest model in it, at $0.055 for 40 turns against $0.080 for deepseek-flash even off-peak, because Haiku's $0.10 input and $0.50 output list below DeepSeek's off-peak $0.15 and $0.60. With caching, deepseek-flash off-peak and Haiku 5.5 come out nearly level ($0.013 and $0.014), and DeepSeek pulls ahead as conversations lengthen, because its $0.003 cache hit is less than a third of Haiku's $0.01 read and it charges nothing to write: at 100 turns the figures are $0.037 and $0.050. V4 Pro at peak, $0.104 cached, still costs less than Gemini 3.8 Flash at $0.111 and well under half of Claude Sonnet 5.5 at $0.240. And the clock matters as much as the model: the same deepseek-flash conversation costs twice as much at 08:00 UTC on a Tuesday as it does at 14:00.

At 10,000 forty-turn conversations a month, deepseek-flash costs $1,596 without caching and $260 with it at peak rates, and $798 and $130 off-peak. On Claude Sonnet 5.5 the same volume costs roughly $10,920 and $2,396.

A conversation's cost also grows faster than its length. Going from 40 to 100 turns multiplies turns by 2.5 and the uncached deepseek-flash bill by more than five, because every turn carries every turn before it. Caching flattens that curve on DeepSeek more than anywhere else, but it has two limits: the cache is best-effort and can be cleared within hours, and it does nothing for a user who comes back tomorrow and needs yesterday's context. The other lever is to stop resending the whole transcript. We build Maximem Synap, a memory layer that returns a ranked set of facts about the user, trimmed to a token budget you set (2,000 tokens by default), in place of the full history. When that beats caching, and when it does not, is worked through in how to reduce LLM token costs in long conversations.

Does DeepSeek give free API credits?

DeepSeek does not publish a free tier or a sign-up credit amount on its pricing page, FAQ or terms as of 27 October 2026. What its documentation does describe is a "granted balance": the pricing page says fees are deducted from "your topped-up balance or granted balance, with a preference for using the granted balance first", the balance endpoint reports granted_balance as "the total not expired granted balance", and the Open Platform terms say any "free quota for paid services" is "subject to the information displayed on our product pages and the platform website". Third-party pages that quote a fixed number of free tokens for new accounts are quoting something DeepSeek's own pages do not state today; check the balance endpoint or the platform's billing page after you sign up to see whether your account received a grant.

DeepSeek API billing is prepaid, with no plans or tiers

DeepSeek is prepaid. You top up through PayPal, bank card, Alipay or WeChat Pay, according to the DeepSeek FAQ, and every request draws down the balance at the rates above. When the balance runs out, requests return HTTP 402 "Insufficient Balance" until you add funds, according to the error code page. The FAQ says topped-up balances do not expire and that unused balances can be refunded; the terms add that a refund is issued "in one lump sum for all remaining unspent amounts", after "necessary costs such as handling fees".

There are no plans or usage tiers. The FAQ describes "a unified pricing standard and no tiered plans", and higher concurrency than the default 2,500 (deepseek-flash) or 500 (V4 Pro) is granted on request "at no additional cost". A request that exceeds the limit gets HTTP 429. Concurrency counts at the account level, across all API keys, from the moment a request is sent until its response completes.

DeepSeek counts roughly 0.3 tokens per English character and 0.6 tokens per Chinese character, according to its token usage page, and publishes an offline tokenizer for exact counts.

Search and Fetch cost $0.75 per 1,000 calls; the Files API is free

Sources: pricing page, Search & Fetch guide, vision guide, Files API guide and the Vision-Exp release note.

Search and Fetch bill per successful request, not per token: a request is billable only when it returns HTTP 200 with a non-empty result, so failed requests, empty searches and timeouts cost nothing, and the count parameter does not change the price. Images are resized before inference, which is why a 2,000 by 2,000 image and a 5,000 by 5,000 image cost the same; at the 1,024-token cap an image costs about $0.0003 on deepseek-flash at peak, or $0.31 per thousand images.

Using DeepSeek in Claude Code or Codex: which price you pay

DeepSeek documents integrations for Claude Code, Codex, OpenCode and several other agent tools, and the model name you send decides the bill. On the Anthropic-format endpoint, any model name starting with claude-opus is mapped to deepseek-v4-pro and billed at the V4 Pro price, while claude-sonnet and claude-haiku names map to deepseek-flash, according to the Anthropic API guide. An agent left on an Opus model name therefore pays 4.4 times the Flash input rate without anyone choosing V4 Pro. The same guide lists cache_control as ignored and budget_tokens as ignored under thinking, so set effort with the effort parameter instead.

The other change lands this week. The web search and page fetches that Claude Code, Codex and DeepSeek Harness make through DeepSeek's server-side tools start billing at $0.75 per 1,000 successful calls from 16:00 UTC on 31 October 2026. An agent that runs 2,000 searches a day adds $1.50 a day.

Together and Fireworks charge DeepSeek's peak price around the clock

Per million tokens, read on 27 October 2026 from Together's pricing page and OpenRouter's public endpoint listings.

Together charges a flat rate that matches DeepSeek's peak price around the clock, so it costs the same as DeepSeek's own API during peak hours and twice as much off-peak, with no peak window to plan around. OpenRouter lists more than thirty endpoints for V4.1 Flash 1017, from $0.022 to $0.45 per million input tokens, and many of the cheapest run FP4 or FP8 quantized weights, so a low price can mean a different numeric precision from the model DeepSeek serves. OpenRouter's listing for DeepSeek's own endpoint shows the off-peak rate, with no peak rate. The Amazon Bedrock pricing page lists DeepSeek-R1, V3.1 and V3.2 and none of the V4 family on 27 October 2026; Bedrock renders those prices in the browser, so check the console for the current figure.

Where DeepSeek processes your data, and what the terms say

DeepSeek's privacy policy (last updated 10 February 2026) says: "To provide you with our services, we directly collect, process and store your Personal Data in People's Republic of China." The Open Platform terms (released 22 April 2026, effective 29 April 2026) are governed by "the laws of the People's Republic of China in the mainland", with disputes going to a court with jurisdiction over the registered office of Hangzhou DeepSeek Artificial Intelligence Co., Ltd.

On inputs and outputs, the terms say you keep whatever rights you have in your inputs, DeepSeek assigns its rights in the outputs to you, and you may use both for purposes that include "training other models (such as model distillation)". The privacy policy does not cover personal data of your own end users: when they use an application you built on the platform, you are the controller and owe them your own privacy policy. We did not find a data processing agreement or a zero-retention option on the DeepSeek pages we read, and the FAQ says its cooperation agreement is "a standard framework agreement and cannot be modified".

For a CIO the practical reading is short. Treat any prompt that contains customer personal data as data that DeepSeek's own policy says is processed and stored in China, under terms governed by Chinese law, and put it through legal review before it goes to api.deepseek.com; teams that need US or EU processing run the open weights through a host in that region, at that host's price.

How DeepSeek API prices have moved in 2026

Sources: DeepSeek change log and news pages.

DeepSeek's terms reserve the right to adjust prices "based on business circumstances" and promise notice "through prominent methods such as in-site notifications, website announcements, or emails" when a change "significantly affects your rights". The name deepseek-flash points at whatever the current Flash version is, and it moved to a new checkpoint on 17 October without a price change, so date your evals and re-run them when the change log moves.

How to estimate your monthly DeepSeek bill, and cut it

A request costs (cache-hit tokens times the hit rate) plus (cache-miss tokens times the miss rate) plus (output and reasoning tokens times the output rate), at the peak or off-peak rate for the time the request is made, plus $0.75 per 1,000 successful Search or Fetch calls. For a conversation, input per turn is the system prompt plus every earlier turn, so estimate per conversation as in the table above, split your traffic into peak and off-peak hours, and multiply by conversations per month.

Six levers move a typical bill, listed roughly in order of size:

1. Use deepseek-flash unless V4 Pro wins your evals. V4 Pro costs 4.4 times as much on uncached input and 3.3 times as much on output.
2. Keep the prefix stable. Put the system prompt, tool schemas and fixed documents first and per-turn content last, and watch prompt_cache_hit_tokens; a hit costs 1/50 of a miss on Flash.
3. Move deferrable work outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. There is no Batch API, and off-peak is the same 50% discount.
4. Set effort per task. low for classification and routing, high for daily agent work, and thinking off where you do not need it; reasoning bills as output, and in tool loops it comes back as input.
5. Check the model name your agent tool sends. A claude-opus name on the Anthropic endpoint bills V4 Pro.
6. Stop resending the whole conversation once chats run long, by summarising or retrieving only what the next turn needs.

Frequently asked questions

How much does the DeepSeek API cost per million tokens?

As of 27 October 2026, deepseek-flash (V4.1-Flash-1017) costs $0.15 per million input tokens on a cache miss, $0.003 on a cache hit and $0.60 per million output tokens off-peak, and $0.30, $0.006 and $1.20 during peak hours. deepseek-v4-pro costs $0.66, $0.022 and $1.98 off-peak and $1.32, $0.044 and $3.96 at peak.

What are DeepSeek's peak and off-peak hours?

Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays (09:00 to 12:00 and 14:00 to 18:00 Beijing time, or 06:30 to 09:30 and 11:30 to 15:30 in India). Every other hour, including all weekend hours, is off-peak at half price.

Does DeepSeek have a free tier or free API credits?

DeepSeek's pricing page, FAQ and terms publish no free tier or sign-up credit amount as of 27 October 2026. The platform supports a "granted balance" that is spent before topped-up funds, and the balance endpoint shows whether your account has one.

Is DeepSeek V4 Pro being retired?

DeepSeek V4 Pro is not being retired as of 27 October 2026. DeepSeek announced in September that V4 Pro requests would route to V4.1 Flash from 14 September, then reversed that in its change log, and V4-Pro-0813 is still served and billed at its own rates.

How much did DeepSeek V3.2 cost, and can I still use it?

DeepSeek V3.2 was served through deepseek-chat and deepseek-reasoner, which DeepSeek retired on 24 July 2026, so DeepSeek's own API no longer offers it. Third-party hosts still serve the open weights; on OpenRouter on 27 October 2026, most providers listed it between $0.21 and $0.30 per million input tokens and $0.31 and $0.96 per million output tokens, with a 163,840-token context.

Does DeepSeek charge for cache writes or reasoning tokens?

There is no cache write or storage charge: caching is automatic and a hit simply bills at the lower rate. Reasoning tokens are billed at the output rate, and thinking mode is on by default at high effort.

Is DeepSeek cheaper than Claude or Gemini?

On cached workloads, usually. A 40-turn conversation with caching costs $0.013 on deepseek-flash off-peak and $0.026 at peak, against $0.014 on Claude Haiku 5.5, $0.111 on Gemini 3.8 Flash and $0.240 on Claude Sonnet 5.5, as of 27 October 2026. On uncached list price, Claude Haiku 5.5 at $0.10 input and $0.50 output is now below deepseek-flash.

Where does the DeepSeek API process data?

DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China, and its Open Platform terms are governed by mainland Chinese law. Teams that need processing elsewhere use the open weights through a host in their own region.

Sources

All retrieved 27 October 2026.

• DeepSeek Models & Pricing and the Chinese-language pricing page
• DeepSeek API documentation: context caching, thinking mode, rate limits, token usage, error codes, Chat Completions reference, balance endpoint, Anthropic API, Search & Fetch, vision, Files API, FAQ, change log and news posts of 24 April, 13 August, 21 August, 10 September and 17 October 2026
• DeepSeek Open Platform Terms of Service and DeepSeek Privacy Policy
• Claude API pricing and prompt caching, Anthropic
• Gemini Developer API pricing, Google
• Together AI pricing, OpenRouter model endpoints and Amazon Bedrock pricing

Worked examples are our own arithmetic on those rates.