DeepSeek's price sheet has changed more often this year than its model names, which is saying something. Here is where it stands in mid-November 2026, with the parts that actually move a bill.

The token table

Prices are per one million tokens, in US dollars.

deepseek-flash (V4.1-Flash-1017), 1M context, 384K max output:
cache hit — $0.003 off-peak / $0.006 peak
cache miss — $0.15 off-peak / $0.30 peak
output — $0.60 off-peak / $1.20 peak
concurrency limit — 2,500

deepseek-v4-pro (V4-Pro-0813), 1M context:
cache hit — $0.022 / $0.044
cache miss — $0.66 / $1.32
output — $1.98 / $3.96
concurrency limit — 500

Flash does not support vision; the multimodal path is V4.1-Flash itself, which does. Pro is the reasoning-heavy option, and it costs between four and five times as much for the same tokens.

Peak hours are narrower than people assume

Off-peak is exactly half price, and the window is defined in UTC: 01:00 to 04:00 and 06:00 to 10:00, Monday through Friday, excluding Chinese public holidays. Everything else — nights, weekends, holidays in full — is off-peak.

Convert that to your own day before planning anything. For a team in UTC+8, peak is 09:00-12:00 and 14:00-18:00, which is the entire working day. For a team in California, peak is 17:00-20:00 and 22:00-02:00, which is mostly the evening. The same batch job is cheap in Shanghai at night and expensive in San Francisco at night, because "night" is not the variable — UTC is.

A server room in an office building. Most of a chat bill is memory held open, not compute spent. Photo: Mike Beltzner / Wikimedia Commons, CC BY-SA 2.0
A server room in an office building. Most of a chat bill is memory held open, not compute spent. Photo: Mike Beltzner / Wikimedia Commons, CC BY-SA 2.0

A worked example

Take a support assistant handling 200,000 requests a month. Each request carries a 6,000-token system prompt plus 400 tokens of conversation, and returns 300 tokens. The system prompt is identical every time, so it is a cache hit; the conversation is a miss.

Input, cached: 6,000 × 200,000 = 1.2B tokens at $0.003 per million = $3.60.
Input, uncached: 400 × 200,000 = 80M tokens at $0.15 per million = $12.00.
Output: 300 × 200,000 = 60M tokens at $0.60 per million = $36.00.
Total at off-peak: about $51.60 a month.

Run the same workload entirely at peak and the total is about $103.20. Nothing about the product changed; only the clock did.

Now break the cache. Drop the stable system prompt to the front of the user message, or inject a timestamp into it, and the 1.2B cached tokens become 1.2B uncached ones: $180 instead of $3.60. The single most expensive mistake in DeepSeek billing is not choosing the wrong model. It is invalidating a cache you did not know you had.

The non-token line items

Search, text or image, is $0.75 per thousand calls. Web fetch is the same. File parsing and OCR are $0.225 per thousand pages. Text to speech is $0.45 per ten thousand characters at list, currently half that. Speech to text is $0.116 per audio hour, also halved for now.

These are small until they are not. An agent that searches on every turn adds $0.75 per thousand turns on top of tokens, which is comparable to the token cost itself for short answers. Gating search behind an explicit intent check is the cheapest optimisation available.

When Pro is worth it

Pro costs four to five times Flash for the same tokens, so the question is whether it saves more than it costs. It usually does on multi-step reasoning where Flash fails and retries: a failed run costs the full price and produces nothing. It usually does not on extraction, classification, summarisation and formatting, where Flash is already accurate and Pro just bills more.

A practical rule: route by task, not by preference. Run both on a hundred real tickets, compare cost per successful completion rather than cost per token, and let the number pick.

Two operational notes

Batch and cache warm-up jobs belong in off-peak windows. Schedule them in UTC and put the conversion in one place, not in six cron files.

And check your invoice against the peak definition once. DeepSeek excludes Chinese public holidays, which means a weekday in early October can be billed at half rate while a Saturday in a US holiday week is not. If your finance team reconciles monthly, that line will confuse them eventually.