Free
Token calculator

DeepSeek API pricing: peak hours, cache hits, real bills

DeepSeek's API costs $0.15-$0.30 per 1M input tokens on Flash and $0.66-$1.32 on V4-Pro, half price off-peak. Monthly bills and rivals, November 2026.

By · Updated

DeepSeek sells two API models, priced per million tokens with a peak rate and an off-peak rate that is exactly half. As of November 6, 2026, deepseek-flash (DeepSeek-V4.1-Flash) costs $0.15 to $0.30 per million input tokens and $0.60 to $1.20 per million output tokens, and deepseek-v4-pro costs $0.66 to $1.32 input and $1.98 to $3.96 output. For US users the useful detail is the clock: DeepSeek’s peak hours fall in the US evening and night, so a workload that runs during a normal US working day pays the lower rate.

We haven’t run workloads on the DeepSeek API ourselves. The prices below come from DeepSeek’s Models & Pricing page, read November 6, 2026, and match AskGPU’s model data (checked November 2, 2026). Billing rules come from DeepSeek’s API docs and FAQ, read the same day. The monthly bills are our own arithmetic from those published rates, using token counts we state for each example.

How much does the DeepSeek API cost per million tokens?

DeepSeek lists three prices per model: input that hits the context cache, input that misses it, and output.

Model Input, cache hit Input, cache miss Output
deepseek-flash, off-peak $0.003 $0.15 $0.60
deepseek-flash, peak $0.006 $0.30 $1.20
deepseek-v4-pro, off-peak $0.022 $0.66 $1.98
deepseek-v4-pro, peak $0.044 $1.32 $3.96

US dollars per million tokens. Source: DeepSeek API docs, Models & Pricing, read November 6, 2026.

The two models share most of their specs, per the same page:

deepseek-flash deepseek-v4-pro
Model version DeepSeek-V4.1-Flash-1017 DeepSeek-V4-Pro-0813
Context length 1M tokens 1M tokens
Maximum output 384K tokens 384K tokens
Thinking mode Thinking and non-thinking (thinking is the default) Same
Vision (image input) Yes No
Concurrent requests per account 2,500 500
JSON output, tool calls, Anthropic-format API Yes Yes

There is no long-prompt surcharge on DeepSeek’s page: a 900,000-token prompt is billed at the same per-token rate as a short one. That differs from several rivals in our data. Claude Haiku 5.5 charges more for prompts over 100,000 tokens, GPT-6 Luna and GPT-6.1 Sol for prompts over 272K tokens, and both Grok models for prompts of 200K tokens or more.

Our model page for DeepSeek V4 Pro has more on the model itself.

When are DeepSeek’s peak and off-peak hours in US time?

DeepSeek introduced peak and off-peak pricing on August 13, 2026, effective from 16:00 UTC on August 16, according to its change log. The pricing page defines peak as 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. Every other hour, including weekends and Chinese public holidays, is off-peak.

Converted to US time (our arithmetic; US standard time began November 1, 2026):

US time zone Peak hours (local)
Eastern (EST, UTC-5) 8 pm to 11 pm, Sunday to Thursday; 1 am to 5 am, Monday to Friday
Central (CST, UTC-6) 7 pm to 10 pm, Sunday to Thursday; midnight to 4 am, Monday to Friday
Mountain (MST, UTC-7) 6 pm to 9 pm, Sunday to Thursday; 11 pm to 3 am, nights from Sunday to Thursday
Pacific (PST, UTC-8) 5 pm to 8 pm, Sunday to Thursday; 10 pm to 2 am, nights from Sunday to Thursday

During US daylight saving time each window moves one hour later in local time. Either way, 9 am to 5 pm on a weekday is off-peak in every mainland US time zone, and so is all of Saturday.

Peak covers 35 of the 168 hours in a week (7 hours × 5 days), about 21%. If your traffic were spread evenly around the clock, your average rate would be (35 × 1 + 133 × 0.5) ÷ 168 = about 0.60 of the peak price. A US business app used mostly in office hours would pay close to the off-peak rate. A consumer app busiest in the US evening would pay more, because that is when the 01:00-04:00 UTC window lands.

DeepSeek bills by when the request is processed. It doesn’t publish a way to queue work for off-peak hours, so if you want the lower rate for a large job, your own scheduler has to start it at the right time.

How does DeepSeek’s context caching cut the bill?

DeepSeek’s caching works differently from Anthropic’s or OpenAI’s, and it is the reason the cache-hit price is so low ($0.003 per million tokens on Flash off-peak, 2% of the cache-miss price).

According to DeepSeek’s context-caching guide:

  • It is on by default for all users. No code change is needed.
  • When a new request starts with the same content (the same prefix) as an earlier one, the repeated part is a cache hit. The match has to cover a stored “cache prefix unit” in full.
  • DeepSeek lists no separate charge for writing to the cache. The pricing page has only hit and miss prices.
  • Each response reports prompt_cache_hit_tokens and prompt_cache_miss_tokens, so you can see what you paid for.
  • It is “best-effort”: DeepSeek does “not guarantee a 100% cache hit rate.”
  • Unused caches are cleared “usually within a few hours to a few days.” The page gives no fixed lifetime.

In practice, keep the parts that repeat (system prompt, instructions, reference documents) at the start of every request, and put the part that changes at the end. In DeepSeek’s own long-document example, the first two requests about the same report both miss; only the third hits.

Caching only helps with input. Output is always charged in full, so a model that writes long answers stays the biggest line on the bill.

What does a month of DeepSeek API use cost? Four examples

These are hypothetical workloads. The token counts are our assumptions; the rates are DeepSeek’s.

Example 1: a support bot, 20,000 replies a month. Each request sends a 3,000-token system prompt plus 500 tokens of conversation, and gets a 300-token reply. That is 70M input tokens and 6M output tokens a month.

Without any cache hits, on deepseek-flash off-peak: 70 × $0.15 + 6 × $0.60 = $10.50 + $3.60 = $14.10.

Now assume 90% of the system-prompt tokens hit the cache (54M hit tokens) and the rest are misses (6M system-prompt tokens plus 10M conversation tokens = 16M):

Model and hours Cache hits Cache misses Output Monthly total
Flash, off-peak 54 × $0.003 = $0.16 16 × $0.15 = $2.40 6 × $0.60 = $3.60 $6.16
Flash, peak 54 × $0.006 = $0.32 16 × $0.30 = $4.80 6 × $1.20 = $7.20 $12.32
V4-Pro, off-peak 54 × $0.022 = $1.19 16 × $0.66 = $10.56 6 × $1.98 = $11.88 $23.63
V4-Pro, peak 54 × $0.044 = $2.38 16 × $1.32 = $21.12 6 × $3.96 = $23.76 $47.26

With traffic spread evenly across the week (the 0.60 factor above), the cached bot costs about $7.45 a month on Flash and about $28.55 on V4-Pro.

Example 2: one developer using a coding agent. Assume 40M input tokens a month, of which 36M are cache hits and 4M are misses, plus 1.5M output tokens.

  • Flash, off-peak: 36 × $0.003 + 4 × $0.15 + 1.5 × $0.60 = $0.11 + $0.60 + $0.90 = $1.61
  • V4-Pro, off-peak: 36 × $0.022 + 4 × $0.66 + 1.5 × $1.98 = $0.79 + $2.64 + $2.97 = $6.40
  • V4-Pro, peak: 36 × $0.044 + 4 × $1.32 + 1.5 × $3.96 = $1.58 + $5.28 + $5.94 = $12.80

Example 3: summarizing 2,000 documents in one off-peak run. Each document is 20,000 tokens with a 1,000-token summary: 40M input and 2M output, no cache hits.

Model Off-peak Peak
deepseek-flash 40 × $0.15 + 2 × $0.60 = $7.20 $14.40
deepseek-v4-pro 40 × $0.66 + 2 × $1.98 = $30.36 $60.72

DeepSeek’s pricing page lists no batch API or batch discount. Starting a job like this off-peak is the nearest equivalent: the same 50% saving, with no wait for results.

Example 4: how far does a $10 top-up go? Assume a typical mix of 80% input and 20% output with no cache hits. A million tokens of that mix costs:

Model and hours Cost per 1M tokens of mix Tokens per $10
Flash, off-peak 0.8 × $0.15 + 0.2 × $0.60 = $0.24 about 41.7M
Flash, peak $0.48 about 20.8M
V4-Pro, off-peak 0.8 × $0.66 + 0.2 × $1.98 = $0.924 about 10.8M
V4-Pro, peak $1.848 about 5.4M

DeepSeek’s token guide says one English character is about 0.3 tokens, so 10.8M tokens is roughly 36 million characters of text in and out.

For your own numbers, the AI token cost calculator does this arithmetic for DeepSeek and the other models in our data.

Is DeepSeek cheaper than ChatGPT, Claude and Gemini’s APIs?

Often, but not always. Here is Example 1’s support bot (70M input, 6M output, no cache hits) priced across the API models in AskGPU’s data, using list prices for short prompts (checked November 2, 2026):

Model Input / output per 1M Monthly cost
Claude Haiku 5.5 $0.10 / $0.50 $10.00
GPT-6 Luna $0.10 / $0.50 $10.00
DeepSeek Flash, off-peak $0.15 / $0.60 $14.10
Mistral Small 4 $0.15 / $0.60 $14.10
GPT-5.4 nano $0.20 / $1.25 $21.50
DeepSeek Flash, peak $0.30 / $1.20 $28.20
Gemini 3.5 Flash-Lite $0.30 / $2.50 $36.00
DeepSeek V4-Pro, off-peak $0.66 / $1.98 $58.08
Mistral Large 4 $0.70 / $2.10 $61.60
Gemini 3.8 Flash $0.75 / $3.75 $75.00
GPT-5.4 mini $0.75 / $4.50 $79.50
Grok 4.3 $1.25 / $2.50 $102.50
DeepSeek V4-Pro, peak $1.32 / $3.96 $116.16
Grok 4.7 $2 / $6 $176.00
Claude Sonnet 5.5 $2 / $10 $200.00
GPT-6.1 Sol $2 / $10 $200.00
Gemini 4 Argon $2 / $10 $200.00
Claude Opus 5.5 $4 / $20 $400.00

Two things stand out:

  • DeepSeek Flash is not the cheapest small model on list price. Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 input and $0.50 output for shorter prompts, below Flash even off-peak. Flash’s edge is elsewhere: no long-prompt surcharge across its 1M context, and a cache-hit price of $0.003 to $0.006 against $0.01 for Haiku 5.5 and Luna in our data.
  • DeepSeek V4-Pro costs far less than the $2-and-up models. Off-peak, it costs about 29% of Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 4 Argon for this workload ($58.08 vs $200).

With caching, the gap to the frontier models widens. Example 2’s coding agent (4M uncached input, 36M cached reads, 1.5M output) costs:

Model Monthly cost
DeepSeek Flash, off-peak $1.61
Claude Haiku 5.5 or GPT-6 Luna $1.51
DeepSeek V4-Pro, off-peak $6.40
DeepSeek V4-Pro, peak $12.80
GPT-6.1 Sol or Gemini 4 Argon $26.60
Claude Sonnet 5.5 (with 4M tokens written to cache at $2.50) $28.60
Grok 4.7 $35.00
Claude Opus 5.5 (with 4M tokens written to cache at $5) $57.20

Cache-write charges are counted only where our data lists them (Claude Sonnet 5.5 and Opus 5.5). Our data also notes that Google bills cache storage separately for its Flash models, which isn’t included here.

Price is only part of the choice. These models differ in quality, speed and features (Flash takes images, V4-Pro doesn’t), and we haven’t measured any of that. Our Claude API pricing guide and the AI models hub cover the alternatives.

Two of the rivals have announced price rises in our data: Gemini 4 Argon’s price doubles to $4 / $20 from December 1, 2026, and Gemini 3.8 Flash’s to $1.50 / $7.50 from January 1, 2027.

Is the DeepSeek API free?

No. The pricing page lists no free tier or free monthly allowance; every token is charged. The page does mention a “granted balance”: charges come from your topped-up balance or a granted balance, and “granted balance used first” when you have both. DeepSeek doesn’t say on that page how or when it grants balance.

What DeepSeek’s FAQ does say about paying:

  • Prepaid only. You top up first, and usage is deducted from the balance. The FAQ lists PayPal, bank card, Alipay and WeChat Pay on the Top Up page, plus corporate bank transfer.
  • No expiry. “Top-up balances do not expire.”
  • Refunds. “Unused balances can be refunded.” Online payments can be refunded from the Billing page; bank transfers need a support ticket.
  • No tiers. “There is a unified pricing standard and no tiered plans.” Higher rate limits are available on request “at no additional cost.”
  • Invoices can be requested from the Billing page.

The FAQ doesn’t state a minimum top-up.

What else can add to a DeepSeek bill?

Item Price or rule
Search API $0.75 per 1,000 successful requests
Fetch API $0.75 per 1,000 successful requests
Thinking mode On by default at “high” effort; it can be switched off or set to low or max
Concurrency 2,500 requests at once on Flash, 500 on V4-Pro; above that you get HTTP 429 errors
Long waits If a request hasn’t started inference after 10 minutes, the server closes the connection

The Search and Fetch prices add up for agents. An agent making 10,000 searches a month adds 10 × $0.75 = $7.50, more than Example 1’s whole cached Flash bill.

Thinking mode matters too. DeepSeek returns the model’s reasoning in a separate reasoning_content field, and the thinking-mode guide doesn’t say how those tokens are billed. Pricing is tokens × price for every token the API reports in usage. For short, simple tasks such as classification, test with thinking turned off ("thinking": {"type": "disabled"}) and compare the usage figures.

What to watch for

  • Peak-hour surprises. A US evening launch, a nightly job at 2 am Eastern, or a team in Asia can land in peak hours and double the rate. Check the hour in UTC, not local time.
  • Prices can change without notice. The pricing page says DeepSeek may adjust prices and recommends checking the page regularly. Its change log shows three pricing events since August 2026: peak pricing in August, a price cut with V4.1 Flash on September 10, and unchanged prices with the October 17 Flash update.
  • Model names move. deepseek-flash now points to DeepSeek-V4.1-Flash-1017. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work and are served by the same model at the Flash price. The original deepseek-chat and deepseek-reasoner names were scheduled for discontinuation on July 24, 2026, per the change log.
  • Caching is not guaranteed. Budget as if some cached input will miss. Example 1 at 90% hits costs $6.16 off-peak on Flash; at 0% it costs $14.10.
  • Deleting your account forfeits the balance. DeepSeek’s FAQ says “any unused balance in your developer account will be forfeited upon deletion.” Ask for a refund first.
  • Logins and region. The FAQ says Google login isn’t supported from mainland China IPs, and WeChat login works only from them. If your email domain isn’t accepted at sign-up, it recommends a major provider such as Gmail or Outlook.

If you would rather run open models on your own hardware than pay per token, our self-hosted AI guide covers what that takes.

Which DeepSeek model should you choose?

  • High-volume, simple tasks (classification, extraction, routing): deepseek-flash, off-peak where you can, thinking off. Compare it with Claude Haiku 5.5 and GPT-6 Luna, which list lower for prompts under 100,000 and 272K tokens respectively.
  • Long documents and big contexts: deepseek-flash. Its 1M context has no long-prompt surcharge, and repeated documents become cache hits at $0.003-$0.006 per million.
  • Coding agents and harder reasoning on a budget: deepseek-v4-pro. At $0.66 / $1.98 off-peak it costs far less than the $2-input models in our data, such as Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon.
  • Image input: deepseek-flash. V4-Pro has no vision.
  • US business-hours apps: either model at close to the off-peak rate, because 9 am to 5 pm in every mainland US time zone falls outside DeepSeek’s peak windows.
  • Apps busiest in the US evening: budget at the peak rate, or move background work to daytime.
  • Teams that need a contract or vendor onboarding: DeepSeek offers a standard framework agreement on request, and its FAQ says it “cannot be modified.”

Sources