DeepSeek API Pricing Explained: What a Chatbot Actually Costs in 2026
DeepSeek API pricing is simple on the surface and easy to misread in practice: two models, three token rates each, and every rate doubles during weekday peak hours. Off-peak, DeepSeek V4.1 Flash costs $0.15 per million input tokens, $0.003 per million cached input tokens, and $0.60 per million output tokens. V4 Pro costs $0.66, $0.022, and $1.98 on the same basis. If you build a chatbot on Flash and most of your traffic lands outside peak hours, 10,000 typical support chats cost roughly $3.32 a month. This article breaks down the full rate card, the peak-hour trap, and the worked math for a real chatbot workload.
The DeepSeek Rate Card (October 2026)
DeepSeek's API currently offers two model IDs. You call them as deepseek-flash and deepseek-v4-pro, and both run on a 1-million-token context window with up to 384,000 tokens of output per request. Flash is served by DeepSeek-V4.1-Flash-1017 and Pro by DeepSeek-V4-Pro-0813, according to DeepSeek's own Models & Pricing page.
| Rate (per 1M tokens, USD) | Flash off-peak | Flash peak | Pro off-peak | Pro peak |
|---|---|---|---|---|
| Input, cache miss | $0.15 | $0.30 | $0.66 | $1.32 |
| Input, cache hit | $0.003 | $0.006 | $0.022 | $0.044 |
| Output | $0.60 | $1.20 | $1.98 | $3.96 |
Three details in that table matter more than the headline numbers. First, cached input is almost free: a cache hit on Flash costs 2% of a fresh input token. Second, output is the expensive line — four times the fresh input rate on Flash, three times on Pro — so long, chatty answers dominate any DeepSeek bill. Third, the two models are not interchangeable on features: Flash accepts image input, Pro does not, and neither model generates images, video, or audio.
Peak Hours: The Rule That Doubles Your Bill
Off-peak rates are exactly half of peak rates, and peak hours are narrow: 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Everything else — including all weekends and Chinese public holidays in full — is off-peak.
For most developers this is good news disguised as a warning. Those peak windows are Beijing working hours, so a chatbot serving users in Pakistan, Europe, or the Americas will see most of its traffic at off-peak rates. But if you run scheduled batch jobs — nightly summaries, bulk classification, backfill pipelines — check the clock first: a job that runs at 08:00 UTC on a Tuesday bills at double the rate of the same job at 12:00 UTC. Schedule batch work outside the two peak windows and you cut that workload's cost in half with zero code changes. This is the same discipline behind our Batch API vs Real-Time analysis: when a job can wait, timing is a discount.
Worked Example: What 10,000 Chats Actually Cost
Take a typical customer-support chatbot. Each user message, with conversation history and a system prompt, sends about 1,500 input tokens: 1,000 fresh tokens and 500 repeated prefix tokens (the system prompt and earlier turns) that hit the cache. Each reply is about 300 output tokens. That is a realistic, slightly generous shape — real support chats are often shorter.
On Flash at off-peak rates, one chat costs: 1,000 fresh input tokens at $0.15 per million ($0.00015), plus 500 cached tokens at $0.003 per million ($0.0000015), plus 300 output tokens at $0.60 per million ($0.00018) — about $0.00033 per chat. At 10,000 chats a month, that is $3.32. Run the same traffic at peak rates and it doubles to $6.63.
On Pro, the same workload costs $0.00127 per chat off-peak — $12.65 a month for 10,000 chats — and $25.30 at peak rates. Pro is roughly four times Flash's price, so the model choice is the single biggest lever on the bill. Use Flash for the routine chats and reserve Pro for the tickets that genuinely need deeper reasoning, and the blended bill lands much closer to the Flash number than the Pro number. To sanity-check your own token mix, run it through our AI API Cost Calculator before you commit to a model.
Scale it up and the picture stays friendly: 100,000 Flash chats a month cost about $33 off-peak, and 1 million chats about $332. A $10 top-up buys roughly 30,000 Flash chats at off-peak rates. That is why DeepSeek shows up at the top of every budget ranking, including our Cheapest LLM API in 2026 comparison — but the ranking only holds if you respect the next two rules.
Two Rules That Decide Your Real Price
Rule 1: Cache your prefix. DeepSeek caches repeated prompt prefixes automatically — no headers, no SDK changes. A long system prompt, a product manual, or few-shot examples are billed at the cache-hit rate after the first request. In the example above, the 500 cached tokens cost almost nothing; without caching, the same 500 tokens at the fresh rate would add $0.000075 per chat, about 23% more per message. The effect grows with prompt length: the longer your system prompt, the more caching matters. Our prompt caching guide explains how to structure prompts so the repeatable part stays identical across requests.
Rule 2: Watch thinking mode. Both DeepSeek models run in thinking mode by default, and the reasoning tokens generated before the answer are billed as output tokens — the most expensive line on the rate card. For a support chatbot, that reasoning is usually wasted money: the user wants a short answer, not a visible chain of thought. Switching to non-thinking mode for routine chats removes that output-token overhead entirely. If a ticket genuinely needs multi-step reasoning, route just that ticket to Pro with thinking on. This is the same pattern behind why AI agents get so expensive: reasoning output quietly becomes the bill.
Search and Fetch: The Small Line Item
DeepSeek also sells standalone Search and Fetch APIs at $0.75 per 1,000 successful requests each. If your chatbot grounds answers in web search, 10,000 grounded chats add $7.50 — more than doubling the $3.32 Flash chat bill above. Note the timing: when you use DeepSeek through DeepSeek Harness, Claude Code, or Codex, server-side tool calls already hit these APIs, but per-request billing for those tool calls only starts at 00:00 Beijing time on November 1, 2026. If you use those tools, check your November bill for a new line item that was free in October.
What DeepSeek Does Not Give You
There is no permanent free tier for the API. DeepSeek's docs mention a granted balance that is spent before your topped-up balance, but they publish no amount, no eligibility rule, and no expiry — so budget on the pay-per-token rates and treat any granted credit as a bonus. The API is prepaid: when the balance runs out, requests fail. Also note the retired names: the old deepseek-chat and deepseek-reasoner model IDs were retired in July 2026. If an old tutorial or config file still uses them, the calls will fail — update to deepseek-flash and deepseek-v4-pro. And if you are migrating from an older setup, the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but are served by V4.1 Flash and billed at the Flash price.
The Verdict
For a chatbot, DeepSeek's pricing is genuinely cheap — but only at off-peak rates, with caching working, and with thinking mode off for routine traffic. Budget Flash at $0.15/$0.003/$0.60 per million tokens, Pro at $0.66/$0.022/$1.98, double both during weekday peak hours, and remember that output tokens — especially reasoning tokens — are where the money goes. Get those four things right and a 10,000-chat support bot costs less than a cup of coffee a month. Get them wrong and the same bot quietly costs four to eight times more, which is exactly how developers end up asking why their API bill is so high.
FAQ
How much does the DeepSeek API cost?
Off-peak, DeepSeek V4.1 Flash costs $0.15 per million input tokens (cache miss), $0.003 per million cached input tokens, and $0.60 per million output tokens. V4 Pro costs $0.66, $0.022, and $1.98. During weekday peak hours (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays) every rate doubles.
Is the DeepSeek API free?
No. The chat app at chat.deepseek.com is free, but the API is prepaid and billed per token. DeepSeek mentions a granted balance spent before topped-up funds, but publishes no amount or eligibility rule, so do not budget around it.
Which DeepSeek model should I use for a chatbot?
Flash for routine chats: it is roughly four times cheaper than Pro, accepts image input, and handles everyday support, summarization, and extraction work. Reserve Pro for tickets that need complex reasoning, and switch thinking mode off for routine chats so reasoning tokens do not inflate your output bill.
Do deepseek-chat and deepseek-reasoner still work?
No. Those model IDs were retired in July 2026. Use deepseek-flash and deepseek-v4-pro instead. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but are served by V4.1 Flash at Flash prices.
When are DeepSeek's peak hours?
Monday to Friday, 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. All other hours, including weekends and Chinese public holidays in full, are off-peak at half price.
How much does DeepSeek web search cost?
The standalone Search and Fetch APIs cost $0.75 per 1,000 successful requests each. Tool calls made through DeepSeek Harness, Claude Code, or Codex are currently free of that charge, but per-request billing starts at 00:00 Beijing time on November 1, 2026.