DeepSeek V4.1 Pricing: One Shipped Row, One Missing Row, and a Cache Line That Changes Both

There is no single "DeepSeek V4.1" price. The line currently has one shipped model and one named tier with no rate at all, and the two sit in the same family page while behaving nothing alike. DeepSeek V4.1 Flash is priced, callable and cheap: $0.15 per million input tokens and $0.60 per million output off-peak, with a cache-hit row at $0.003. DeepSeek V4.1 Pro has never had a price published — not an introductory rate, not a ceiling, not a rate-card dimension — because it has never had a model card, an endpoint or a repository. If you searched for V4.1 pricing and expected a table, the honest answer is that the table has one row you can pay and one row you can only estimate.

The row you can pay

Read from DeepSeek's Models & Pricing page on 30 October 2026, per million tokens, off-peak with peak in brackets:

• Input, cache hit — $0.003 ($0.006).

• Input, cache miss — $0.15 ($0.30).

• Output — $0.60 ($1.20).

• Peak windows — 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Off-peak is exactly half.

• Limits — 1M-token context, 384K maximum output, 2,500 concurrent requests, text and image input.

• What you actually call — the identifier code>deepseek-flash/code>, which DeepSeek's documentation says is served by DeepSeek-V4.1-Flash-1017. The legacy names code>deepseek-v4-flash/code> and code>deepseek-v4-flash-vision-exp/code> are still accepted, are served by the same checkpoint, and are billed at the Flash price.

That cache line is 2% of the miss price, which makes it the single largest lever on a V4.1 bill — larger than the peak doubling, and larger than any comparison against another vendor. A workload that re-sends a stable prefix is paying three tenths of a cent per million for the part it repeats.

The row you cannot pay

DeepSeek V4.1 Pro is a name used once, in the 10 September release note that shipped Flash, in a sentence about routing Pro requests to Flash "until V4.1-Pro launches." That plan was withdrawn after developers objected to a production model being swapped underneath them. Nothing has been published since: no price, no card, no identifier, no benchmark, and no repository.

So any V4.1 Pro price you see quoted is one of two things, and neither is a vendor number. The first is DeepSeek V4 Pro's card — $0.66 input, $1.98 output, $0.022 cached — used as a stand-in for what a Pro-tier V4.1 model would plausibly cost. The second is a guess. The difference matters: the first is labelled and defensible, the second is how a budget ends up describing a product that does not exist.

What a task costs on the row that exists

A document pass with one million input tokens — 800,000 of them a reused prefix — and 100,000 output tokens:

• Cached, off-peak — 0.8M at $0.003 is $0.0024; 0.2M at $0.15 is $0.03; 0.1M of output at $0.60 is $0.06. Total: under ten cents.

• Uncached, off-peak — $0.15 plus $0.06: twenty-one cents.

• Uncached, in a peak window — $0.30 plus $0.12: forty-two cents.

The same request ranges from about $0.09 to $0.42 without changing a token. Caching accounts for most of that spread, scheduling for the rest, and the model itself for none of it — which is the useful thing to know before comparing V4.1 Flash against anything else on price.

Why the line is priced this way

V4.1 Flash activates 8 billion parameters during prefill and 16 billion during decode on a 552-billion-parameter Causal Encoder-Decoder architecture. The price reflects that: reading a 1M-token context is cheap because most of the network is not computing it. The Pro tier in the V4 generation activated 49 billion of 1.6 trillion, which is why V4 Pro's card is roughly four and a half times Flash's.

DeepSeek's own release note says V4.1 Flash comprehensively surpasses the previous flagship on performance, cost, speed and total runtime. If a V4.1 Pro eventually ships, the pricing question is not whether it costs more than Flash — it will — but whether DeepSeek keeps the family's workhorse pattern or breaks it for the top tier. The company has no history of frontier pricing and a strong history of undercutting, which is the only honest basis for a forecast.

Three things to do before you budget V4.1

• Pin the identifier, not the marketing name. code>deepseek-flash/code> is served by a checkpoint that was released in October 2026. A cost or quality baseline from September describes a model you no longer receive under that name.

• Measure the cache-hit share first. It is the difference between $0.09 and $0.21 on the worked example above, and it is entirely a function of whether your prompt prefix is stable. A timestamp at the top of a system prompt silently converts every call to the full price.

• Budget the peak windows as a real line. Seven weekday hours double every rate. If your users are in Europe or Asia, most of your interactive traffic lands inside them.

DeepSeek V4.1 Flash is on OrcaRouter at DeepSeek's published rate — $0.15 and $0.60 off-peak, $0.003 cached, with the peak doubling applied exactly as the vendor defines it — passed through with no markup, on the same key as a 224-model catalogue, with automatic failover across provider paths. V4.1 Pro is not there and cannot be until DeepSeek publishes an identifier; the checks that would change that are a changelog entry, a repository, or a third column on the price page, and none has moved.