DeepSeek API pricing: V4 Pro, V4.1 Flash, cache hits and off-peak hours

The DeepSeek API has 2 models. DeepSeek V4.1 Flash (deepseek-flash) costs $0.30 per million input tokens and $1.20 per million output at peak hours. DeepSeek V4 Pro (deepseek-v4-pro) costs $1.32 and $3.96. Those headline prices matter less than 2 discounts that apply with nothing to switch on. Input that hits DeepSeek's cache costs 2% to 3% of the normal price, and every request outside peak hours costs half.

The full price list

Per million tokens, in US dollars:

Both models take up to 1 million tokens of context and write up to 384,000. Flash also reads images; Pro does not. A web search tool costs $0.75 per 1,000 successful requests, billed per request from 1 November 2026. The older model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are now served by Flash, at Flash's price.

Peak hours, in UTC

Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. Everything else, including all weekend, is off-peak. Those are Beijing working hours (09:00 to 12:00 and 14:00 to 18:00). The second window is a European morning, so a European team should schedule batch work from the afternoon on; for teams in the Americas, almost the whole working day is off-peak.

The cache does most of the work

DeepSeek caches the start of every prompt on disk automatically. When a later request begins with the same tokens, those tokens are billed at the hit price. The response's usage reports prompt_cache_hit_tokens and prompt_cache_miss_tokens, so you can see your hit rate on every call.

Slide the share of input that hits the cache and watch the bill:

Try itWhat a month on the DeepSeek API costs

Job

·Support chatbot·Coding agent·Overnight labels

V4.1 FlashV4 Pro

PeakOff-peak

Input from the cache

80%

$86.88

Input, cache hit	$2.88
Input, cache miss	$36.00
Output	$48

600M input and 40M output tokens a month.

The same rules as any prompt cachePrompt CachingLetting the API keep the processed start of a prompt for a few minutes, so later requests that begin the same way are cheaper and quicker. decide your hit rate: keep the system prompt, tools and examples first and identical, and put what changes at the end. DeepSeek calls its cache best effort, and entries are cleared after hours to days without use, so a nightly job that runs once may not find yesterday's prefix.

3 worked monthly bills

1. A support chatbot on Flash, at peak. 100,000 replies a month, each sending 6,000 tokens (system prompt plus conversation so far) and writing 400. The conversation grows turn by turn, so about 80% of input hits the cache.

Output is more than half the bill. Shorter answers save more here than a better cache.

2. A coding agent on Pro. 300,000 model calls a month, each sending 40,000 tokens of instructions, files and history and writing 1,500, with 90% of input cached.

With no cache hits, the input alone would cost $15,840 and the month $17,622. An agent's long, repeated context is exactly what the cache is for, which makes a change that breaks the prefix the most expensive bug in this design.

3. Overnight classification on Flash, off-peak. 2 million documents, 1,500 tokens each, 50 tokens of label out. Only the short shared instruction repeats, so about 10% of input hits.

Run the same job during peak hours and it costs $931.80. Here the clock saves more than the cache.

What else to know before you switch

• Concurrency is capped: 2,500 requests at once on Flash and 500 on Pro. A large batch job needs a queue, and rate limits explains the retry side.
• Credits are spent in order: granted (promotional) balance is used before money you topped up.
• Price is per token, value is per answer. A cheaper model that needs a second try or more human review can cost more per finished task. Run 50 of your real requests on each candidate before comparing bills.

Our DeepSeek V4.1 Flash price page and V4 Pro page keep the current prices, and the cost calculator prices the same job on DeepSeek, Claude, GPT and Gemini side by side.