DeepSeek API Pricing in 2026: Flash vs Pro, Peak & Costs

DeepSeek’s direct API bills $0.15 per million input tokens and $0.60 per million output tokens for Flash off-peak; Pro is $0.66 input and $1.98 output. Peak rates (01:00-04:00 and 06:00-10:00 UTC, weekdays, excluding Chinese public holidays) are double. Cached input is 50x cheaper than cache-miss input. For a typical 60-messages-a-day Janitor AI chat at 8k context with half a swipe each, Flash costs about $5 a month off-peak and Pro about $21.

DeepSeek’s pricing page is short, but three details change the bill more than people expect: peak hours, cache hits, and the fact that every swipe resends the whole context. This is what the numbers actually look like for roleplay use.

Key takeaways

• Flash off-peak: $0.15 input / $0.60 output / $0.003 cached input, per million tokens
• Pro off-peak: $0.66 input / $1.98 output / $0.022 cached input, per million tokens
• Peak doubles the rate; peak hours are two UTC windows on weekdays only
• Cache hits can make a hot chat cost 10x less than a cold one
• Prepaid balance: plan the top-up at the model and context you actually use

---

The full DeepSeek API price table

Prices are USD per million tokens. The table below is DeepSeek’s live pricing page, checked 2026-11-04 at api-docs.deepseek.com/quick_start/pricing.

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

Both models expose a 1M-token context window and up to 384k output tokens, and both support thinking and non-thinking modes. Flash is DeepSeek-V4.1-Flash-1017; Pro is DeepSeek-V4-Pro-0813. The old deepseek-v4-flash and deepseek-v4-flash-vision-exp IDs are still accepted but now served by Flash at the Flash price.

When are DeepSeek’s peak hours?

DeepSeek’s note on the pricing page is precise:

Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full.

That is seven hours a day across two windows, five days a week. The rest of the week, more than 80% of all hours, is off-peak. If you chat in the US evening or European afternoon, you are almost entirely off-peak. If you chat during the European morning or Indian afternoon, you are at least partially inside a peak window.

Practical read: unless you run batch work during Chinese business hours, treat the headline rates ($0.15 / $0.60 for Flash) as what you will actually pay. The peak rates mostly matter for multi-region production apps, not evening roleplay.

What is cache hit input?

DeepSeek caches recent prompts and bills repeat prompt prefixes at the cache-hit rate: Flash’s $0.003 and Pro’s $0.022 are 50x and 30x cheaper than their respective cache-miss input prices. In a Janitor AI or SillyTavern chat, the character card, persona and lorebook that go in front of every message are stable; if you keep the same chat open, the second and later messages hit the cache for most of their input.

A realistic mix for a running roleplay session is 60-90% cache hits on the character card and recent history, miss on the new user turn. The calculator at our proxy cost tool lets you enter a cache-hit share and shows the effect.

A rough shape: if 70% of your 8k context is cache-hit on Flash off-peak, the input side drops from $0.00120 per request to $0.00037, so a $5-a-month bill becomes closer to $2.

Cost per message and per month

Using DeepSeek’s own prices, 30% peak share (typical in Europe and North America combined), 0% cache hits (worst case, cold chat start), 8k-token context, 350-token reply, and 0.5 swipes per message:

Model	Per message (incl. swipes)	Per day at 60 msgs	Per month
Flash	$0.0027	$0.16	about $5
Pro	$0.0116	$0.70	about $21

Lighter use, 10 messages a day at 4k context with no swipes, no peak, no cache:

Model	Per message	Per day	Per month
Flash	$0.00088	$0.009	about $0.27
Pro	$0.00374	$0.037	about $1.12

Heavy use, 150 messages a day at 16k context with one swipe each:

Model	Per message	Per day	Per month
Flash	$0.0079	$1.19	about $35
Pro	$0.0339	$5.08	about $150

Run your own numbers in the proxy cost calculator; the maths above comes from the same formula.

DeepSeek direct vs Janitor+ vs NanoGPT

The direct DeepSeek API competes with two flat-price alternatives for roleplay:

Route	Flat cost	Variable cost	Best when
DeepSeek Flash direct	$0	$0.15 input / $0.60 output per 1M tokens (off-peak)	You chat 30-120 messages a day and want the lowest paid token price
Janitor+	$12.99/month or $99/year	Router usage per token on top	You prefer one Janitor subscription and like the janitor+ Router UI
NanoGPT subscription	$12/month	Pay-as-you-go above plan	You use open-source text models outside Janitor too

Break-even math: Janitor+ at $12.99/month covers its own JLLM+ model. If you want a DeepSeek model specifically, Janitor+ Router bills you per token on top, roughly at provider rates. A direct DeepSeek Flash key costs about $5/month at 60 messages/day; Janitor+ costs $12.99/month flat before any Router usage. Direct wins at light and medium use. Janitor+ wins when you also value the JLLM+ model, priority access and more context.

Your Router model and context will shift these numbers. The Janitor AI proxy guide covers each route’s privacy trade-off; the DeepSeek on Janitor AI setup guide covers the direct-API steps.

Getting the prepaid balance right

DeepSeek bills from a prepaid balance. For a first top-up:

• $2 covers a typical Flash user for two or three months
• $10 covers a heavy Flash user for one to two months, or a Pro user for about three weeks
• Avoid loading more than one or two months of usage on a new key; rotate the key first if it has been in a proxy config

Keep your DeepSeek balance, OpenRouter balance and Janitor+ Router wallet separate; pasting one key against another’s endpoint is the most common 404 and 402 you will see.

What to watch for

• Model aliases. deepseek-v4-flash still works but is served by Flash at the Flash price. There is no reason to keep using the old alias; it adds confusion when reading billing logs.
• Thinking mode. Both Flash and Pro default to thinking mode. Thinking tokens are output tokens and are billed at the output rate; the extra tokens are often a third of a short reply’s billed output.
• Search and Fetch APIs. DeepSeek added paid Search and Fetch APIs: $0.75 per 1,000 successful requests from 1 November 2026. Normal chat requests do not touch these; they only matter if you call the Search/Fetch endpoints or use DeepSeek Harness/Claude Code/Codex server-side tools.
• Concurrency limits. Flash allows 2,500 concurrent requests, Pro 500. Single-user roleplay never hits either.

Sources & references

• DeepSeek API docs, Models & Pricing: the current Flash and Pro rates, peak hours, cache-hit and cache-miss definitions, concurrency limits, and the retired-alias note. Checked 2026-11-04. api-docs.deepseek.com/quick_start/pricing
• OpenRouter API docs, API Credit & Rate Limits: free-model cap, 10-credit unlock and 9-credit grant, used for the direct-vs-OpenRouter comparison. Checked 2026-11-04. openrouter.ai/docs/api-reference/limits
• Janitor AI subscription page: Janitor+ at $12.99/month or $99/year. Used for the break-even math. Checked 2026-11-04. janitorai.com/plus
• NanoGPT pricing page: the $12/month open-source text-model subscription. Checked 2026-11-04. nano-gpt.com/pricing