DeepSeek Pricing (September 2026): V4.1 Flash and V4 Pro API Rates, the Peak Hours US Teams Never Hit, and What a Real Workload Costs

DeepSeek pricing: the short answer

DeepSeek's chat app is free, with no paid tier. The API is pay-per-token, and as of September 17, 2026 DeepSeek's own pricing page lists V4.1 Flash at $0.30 per million input tokens and $1.20 per million output tokens during peak hours, and exactly half that, $0.15 and $0.60, off-peak. V4 Pro is $1.32 and $3.96 at peak, $0.66 and $1.98 off-peak. The part the roundups miss: peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which is nighttime across the United States, so a US team running jobs during the workday pays the off-peak rate. Prices dropped again on September 10, when V4.1 Flash shipped. Here is every rate, what changed, what a real workload costs next to OpenAI and Claude, and the cost that is not on the pricing page.

The full DeepSeek API price table

Per million tokens, read from api-docs.deepseek.com on September 17, 2026:

Both models support thinking and non-thinking modes, JSON output, tool calls, OpenAI's Responses API format, and an Anthropic-format endpoint. The bill is tokens times price, taken from a prepaid balance; if you have a granted balance, DeepSeek uses it first.

Peak vs off-peak: why US business hours bill at half price

Since August 16, 2026, DeepSeek has charged two rates. Off-peak is half of peak. Peak is two windows, 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Everything else, including the whole weekend, is off-peak. On US daylight time that works out to:

• Eastern: peak 9pm to midnight and 2am to 6am
• Central: peak 8pm to 11pm and 1am to 5am
• Pacific: peak 6pm to 9pm and 11pm to 3am

So a US business calling the API from 9 to 5 local time pays off-peak rates. The trap runs the other way: the overnight batch job someone scheduled for 2am Eastern because "nobody is using anything then" lands squarely in DeepSeek's peak and costs twice as much. When US clocks fall back on November 1, each local window moves an hour earlier; a West Coast team still working at 5pm Pacific in the winter is in peak.

What changed in 2026

• April 24: V4 launched on the API as V4 Pro and V4 Flash. The old deepseek-chat and deepseek-reasoner names were discontinued on July 24.
• August 13: V4 Pro went generally available.
• August 16: peak and off-peak pricing took effect at 16:00 UTC.
• September 10: V4.1 Flash released with built-in image understanding, and API prices were reduced. V4 Flash and V4 Flash Vision Exp were retired; calls to those old names are served by V4.1 Flash at the Flash price. DeepSeek also said it will keep serving V4 Pro after September 14 with billing unchanged.

If a blog post quotes DeepSeek at a flat rate with no peak hours, it was written before August 16. If it quotes V4 Flash as a current model, it was written before September 10.

Flash or Pro: which one to pay for

Start on Flash. It is about a quarter of Pro's price on input and less than a third on output, it reads images, and it allows five times the concurrent requests. Pay for Pro only after you have a specific task where Flash's answers are measurably wrong and Pro's are not. "Pro sounds better" is how API bills quadruple for no visible difference in the output a customer sees.

What a real workload costs: DeepSeek vs OpenAI vs Claude

A labeled assumption, not a benchmark: a support assistant handling 1,000 conversations a month, averaging 3,000 input tokens (instructions, history, a help article) and 500 output tokens each. That is 3 million input and 500,000 output tokens, with no caching discount applied:

OpenAI and Anthropic rates are from my OpenAI API pricing and Anthropic API pricing breakdowns. Two things jump out. DeepSeek is not automatically cheapest: at peak rates, Flash costs more than GPT-5.6 Luna on this mix. And at small-business volume, every option on the table is under $15 a month. The spread that looks dramatic per million tokens is lunch money per month. Price is rarely the deciding line item for a business at this scale.

The cost that is not on the pricing page

DeepSeek's hosted chat and API run under its own privacy terms, and those terms say personal data is stored in the People's Republic of China and used to improve its models unless you opt out. I went through the policy line by line on is DeepSeek safe. The practical rule I give clients: public material and throwaway tests through DeepSeek's API are fine; customer records, client documents, and anything under a confidentiality agreement are not.

If you want DeepSeek's price on private data, the route is the open weights. They are MIT-licensed, so you can run them on infrastructure you control and pay for hardware instead of tokens. That trade only pays off at real volume, as covered on is DeepSeek free. For most small businesses the honest math is the table above: a vendor whose business terms already say no training costs a few dollars more a month.

What to do today

1. If you call DeepSeek's API, check when your jobs run. Anything scheduled between 01:00 and 10:00 UTC on a weekday may be in peak. Move it to US daytime and the bill halves.
2. Switch any code still calling deepseek-v4-flash to deepseek-flash. The old name works for now, but DeepSeek describes that routing as temporary.
3. Price your workload, not the model. Estimate a month of input and output tokens and run it against three vendors. If every answer is under $20, choose on data terms and output quality, not the per-million rate.

If the real question is not which API is cheapest but which jobs in your business should run on AI at all, and on whose terms, that is what the AI Readiness Assessment answers: $2,500 flat, a written roadmap in two weeks, credited in full if we build it together within 90 days.