Lesson 05 · DeepSeek Mastery Pro ~8 min read Updated September 11, 2026

What DeepSeek costs: about as close to free as AI gets.

DeepSeek's whole pitch is price. Here is the honest breakdown — what is free, what the API actually costs, and the discounts that make a real workload almost free.

01Free where it counts

The chat app and website are free. The API is paid but priced far below the frontier, and new accounts usually get a free starter grant of tokens to build with before you spend anything.

02API pricing, in plain terms

You pay per token, split into input and output, with Flash dramatically cheaper than Pro. The numbers are small enough that for most personal and small-team use, a month of real work costs a few dollars — sometimes cents. Always check DeepSeek's current rate card before relying on a figure.

As of August 13, 2026, the flagship deepseek-v4-pro endpoint — now serving the general-availability V4 Pro 0813 build that left preview on August 13 — lists these rates per million tokens:

V4 Pro list rates, per million tokens

  • Input, cache miss: $0.435
  • Input, cache hit: $0.003625
  • Output: $0.87

Read those first two rows again: a cache hit costs about 1/120th of a cache miss. That is not a rounding difference, it is the whole game — which is why the section below is the one that actually decides your bill. The endpoint carries a 1M-token context window and a 384K-token maximum output.

The discount that matters

Context caching. When your prompts share a repeated prefix (a system prompt, a long document), cached tokens are billed at a fraction of the normal price. For repetitive workloads this alone can cut the bill by most of it.

03The price rise now has a date and a number

The warning that used to sit on DeepSeek's rate card has resolved. Alongside the V4 Pro general-availability release, DeepSeek published a new card that replaces flat pricing with peak and off-peak billing, off-peak set at half the peak rate. It takes effect at 16:00 UTC on August 16, 2026. Peak hours are 01:00–04:00 and 06:00–10:00 UTC; every other hour is off-peak.

V4 Pro from August 16 — per million tokens (off-peak / peak)

  • Input, cache miss: $0.66 / $1.32  (was a flat $0.435)
  • Input, cache hit: $0.022 / $0.044  (was a flat $0.003625)
  • Output: $1.98 / $3.96  (was a flat $0.87)

V4 Flash from August 16 — per million tokens (off-peak / peak)

  • Input, cache miss: $0.22 / $0.44  (was $0.14)
  • Input, cache hit: $0.007 / $0.014  (was $0.0028)
  • Output: $0.66 / $1.32  (was $0.28)

Do the arithmetic before you shrug. V4 Pro output goes up 2.3× off-peak and 4.6× at peak. The cache-hit rate — the one section 02 called the whole game — rises about 6× off-peak and 12× at peak. Caching is still a huge win (a hit is now roughly 1/30th of a miss) but noticeably less huge than the 1/120th it was.

The part that softens it

Peak is only seven hours out of twenty-four, and those hours sit in the small hours of the US morning. If your workload is batch — nightly summarisation, backfills, evals, anything with no human waiting — moving it outside 01:00–04:00 and 06:00–10:00 UTC halves the token bill for the price of editing one cron line. Interactive traffic from US or European users mostly lands off-peak already.

This is still cheap in absolute terms — $1.98 per million output tokens off-peak is far under frontier pricing — but "DeepSeek is nearly free" stopped being a safe planning assumption on August 16. If you sized a budget against the old flat card, re-run it against the peak column, because that is your worst case.
Prices and free grants change. Use the figures here as ballpark and confirm on DeepSeek's pricing page — and model your own volume in our AI cost calculator.

04September 10, 2026: Flash gets cheaper again — and a new name

Three weeks after the peak/off-peak card landed, DeepSeek released DeepSeek-V4.1-Flash and cut the Flash rates with it. The API model name is now simply deepseek-flash; the old deepseek-v4-flash and deepseek-v4-flash-vision-exp names still resolve, but they are served by V4.1 Flash and billed at the Flash price. The rate card as published on the day:

V4.1 Flash from September 10 — per million tokens (off-peak / peak)

  • Input, cache miss: $0.15 / $0.30  (was $0.22 / $0.44)
  • Input, cache hit: $0.003 / $0.006  (was $0.007 / $0.014)
  • Output: $0.60 / $1.20  (was $0.66 / $1.32)

That is roughly a third off input on a miss, more than half off a cache hit, and 10% off output — and the new Flash reads images natively, which V4 Pro does not. The cache-hit ratio is back to about 1/50th of a miss, so the “caching is the whole game” advice in section 02 gets stronger, not weaker. Peak hours are unchanged.

V4 Pro is staying

DeepSeek had signalled that V4 Pro API service would end after September 14, 2026. On September 10 it reversed that: deepseek-v4-pro continues “in response to user demand,” at the same $0.66 / $1.32 input, $0.022 / $0.044 cache-hit and $1.98 / $3.96 output rates listed above. If you had a migration deadline in your calendar for the 14th, delete it — but read Flash vs Pro before you decide Pro is still worth 4× the price, because on DeepSeek’s own benchmarks the new Flash now beats it on most agent and coding work.

Frequently asked

DeepSeek — your questions, answered

How much does DeepSeek cost?
The chat app is free. The API is paid per token but far cheaper than frontier models. Until 16:00 UTC on August 16, 2026 the deepseek-v4-pro endpoint lists $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, and $0.87 per million output tokens. From that moment DeepSeek switches to peak/off-peak billing: V4 Pro becomes $0.66/$1.32 input on a cache miss, $0.022/$0.044 on a cache hit, and $1.98/$3.96 output — off-peak and peak respectively. Verify current rates on DeepSeek's pricing page.
Does DeepSeek have a free tier?
Yes — the chat app and website are free, and the API typically includes a free starter grant of tokens.
Why is DeepSeek so cheap?
Efficient training and serving plus open weights let it undercut frontier pricing dramatically, which is what pressured competitors to lower their prices.
What is context caching?
A discount where repeated prompt prefixes (like a system prompt or a long document) are billed at a fraction of the normal token price — a big saving on repetitive workloads.
Is DeepSeek about to get more expensive?
It already is. DeepSeek published the specifics alongside the V4 Pro general-availability release: from 16:00 UTC on August 16, 2026 the API moves to peak and off-peak billing, off-peak at half the peak rate. V4 Pro output rises from a flat $0.87 per million tokens to $1.98 off-peak and $3.96 at peak — roughly 2.3x and 4.6x.
What are DeepSeek's peak and off-peak hours?
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC. Every other hour is off-peak and billed at half the peak rate. Because peak is only seven hours out of twenty-four, moving batch jobs outside those windows halves their token cost.