DeepSeek API pricing: Flash at $0.30 in, half price off-peak

How much does the DeepSeek API cost per token?

The DeepSeek API sells two models, and each has a peak and an off-peak price. We read every figure below from DeepSeek's Models and Pricing page on 8 November 2026. Prices are US dollars per million tokens:

Flash is the default buy. It costs 4.4 times less than V4 Pro on fresh input and 3.3 times less on output. DeepSeek's own launch note for V4.1 Flash says it's ahead of V4 Pro, too. A cache hit is the cheapest token on the list. When your prompt repeats a prefix DeepSeek has seen, Flash charges 2% of the fresh price.

Both models think by default and can switch thinking off. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work. DeepSeek serves them on V4.1 Flash and bills them at the Flash price.

What does a month of DeepSeek API use cost?

A small app costs $13.68 to $27.36 a month on deepseek-flash. Here's the app that we priced. It gets 1,000 requests a day for 30 days. Each of those requests sends a 2,000-token system prompt that DeepSeek caches, 1,000 fresh tokens and gets 500 tokens back. That's 60 million cache-hit tokens, 30 million fresh ones and 15 million output tokens a month:

Here is the Flash math at peak. 60 × $0.006 is $0.36, 30 × $0.30 is $9.00 and 15 × $1.20 is $18.00. That adds up to $27.36. Off-peak, each line halves, so it's $13.68. On V4 Pro the same month is $101.64 at peak ($2.64 + $39.60 + $59.40) and $50.82 off-peak.

Output is most of the bill. It's two thirds of the Flash total, though it's only a seventh of the tokens. So the cheapest fix is to ask for a shorter answer, and the size of the prompt matters less. DeepSeek's token guide puts "1 English character ≈ 0.3 token," which makes a 500-token answer about 1,700 characters.

When are DeepSeek's off-peak hours?

DeepSeek's off-peak hours are every hour outside 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Weekends and Chinese public holidays are off-peak all day. That's 35 peak hours a week. The other 133, or 79% of the week, bill at half price. Here's the peak window in local time, with US and European clocks already set back for winter:

So if you work in the US, you get the off-peak rate for the whole of your working day. 9 a.m. to 5 p.m. never touches the window in any mainland US time zone. A European team doesn't: its morning from 7 to 11 is peak. If you're in India, you lose most of the workday. DeepSeek's V4.1 Flash note puts it plainly: "Off-peak rates are 50% of peak rates. Schedule flexible workloads off-peak to save."

There's no switch to flip. The rate follows the clock, so a batch job moved from 09:00 UTC to 10:00 UTC pays half. US evening jobs that start between 8 and 11 p.m. Eastern are the ones to move.

Is DeepSeek cheaper than Claude and GPT?

DeepSeek is cheaper than every mid-tier Claude and GPT model, but not the smallest ones. We read OpenAI's API pricing and Anthropic's pricing docs on the same day, 8 November 2026, and priced the same month of app traffic on each:

Against the models that most people compare it with, the gap is wide. Flash costs about an eighth of GPT-6.1 Sol or Claude Sonnet 5.5 at peak, and a sixteenth off-peak. V4 Pro at peak lands next to GPT-6.1 Terra. We left out Anthropic's cache-write charge, which only adds to Claude's totals.

The surprise sits at the bottom. GPT-6 Luna and Claude Haiku 5.5 both list $0.10 in and $0.50 out, below Flash even at night. DeepSeek still wins on cache hits, at $0.003 against $0.01. Luna and Haiku are each lab's smallest tier, so price your own prompts on both before you switch. Kimi pricing runs the same check on Moonshot's K3.1 Pro. Anthropic's Batch API also takes 50% off Claude, which narrows the gap for jobs that can wait.

Which DeepSeek model should you pay for?

Pay for deepseek-flash unless a test on your own prompts shows that V4 Pro does better. DeepSeek says as much itself. Its launch note for V4.1 Flash says "Tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime." Flash also reads images and allows 2,500 requests at once, against 500 on V4 Pro.

V4 Pro nearly went away. The same note said all deepseek-v4-pro requests would route to Flash from 14 September 2026. DeepSeek's changelog then said V4 Pro's API "will continue" past that date, and the pricing page still lists it. If you run V4 Pro in production, keep a Flash fallback ready.

What changed in DeepSeek pricing in 2026?

Four changes in 2026 moved the DeepSeek API bill, and the old figures still circulate. If a page quotes deepseek-chat or V4 Flash prices, it predates the list below:

The September cut came with V4.1 Flash. DeepSeek's note said: "V4.1-Flash lets us serve more users at a lower cost. We're passing the savings on to you." That cut is why price trackers disagree: many still list V4 Flash, which DeepSeek retired that day.

What else does the DeepSeek API charge for?

Web search is the one extra. DeepSeek's Search and Fetch APIs cost $0.75 per 1,000 successful requests each, per its pricing page. DeepSeek bills only successful requests. The Chinese page lists the same fee as ¥5 per 1,000.

You pay for all of it from a prepaid balance. DeepSeek deducts from a granted balance first, then from money you topped up. There's no monthly plan. The chat app costs nothing: DeepSeek's download page says "Chat with leading AI models for free." The API and the app share one account, but only API calls draw on the balance.

What does DeepSeek cost next to the rest of your AI stack?

In a common developer stack, DeepSeek is the cheapest line on the bill. Here's that stack: the app above on Flash at $13.68 to $27.36, plus ChatGPT Plus at $20 a month and Claude Pro at $20 a month for the chat side. That's $53.68 to $67.36 a month, and the two chat plans are 59% to 75% of it.

Whizi, which publishes this page, can only touch the chat part. Whizi Pro costs $29.99 a month and puts Claude, GPT and Gemini in one chat, plus DeepSeek V4.1 Flash at 1 credit a message. It also makes pictures and, on Powerhouse, video and music. It has no API, so your DeepSeek bill stays exactly where it is. If you only use one chatbot, a single $20 plan costs less than Whizi Pro. The cheapest way to use Claude and GPT runs that math for each mix of tools.