DeepSeek Flash API pricing

What it costs on a real workload

Per 1M tokens processed, at the rates above. How the four profiles are defined.

Compare every model on these profiles →

Observed rate history

Dates are when we observed the rate in our own snapshots, which can lag the provider's announcement. Rates per 1M tokens.

What the record shows

We have observed 2 changes to DeepSeek Flash's rates since it entered our archive on 2026-06-22. In the most recent, on 2026-09-14, DeepSeek cut the output rate by 9%, from $1.32 to $1.20 per 1M. DeepSeek renamed this model from deepseek-v4-flash, which we observed on 2026-09-14. Rows dated before then were published under that name. It is one model with two names rather than two models, so the record is continuous and is shown as one.

Against the 40 models we price, it is the 12th cheapest for retrieval-shaped work, which reads a great deal and writes little, and the 12th cheapest for content-shaped work, which does the reverse. The two are close, so it holds its position whichever way your workload leans.

It charges 4.0 times more for output than for input, against a median of 5.0 times across the set. That is narrower than most, so it punishes long answers less than the typical model does.

Inside DeepSeek's own lineup it is the cheapest of the 2 models we track on this workload, which makes it the floor to measure the rest of the house against before looking outside it at all.

7 of the 40 models we price offer a larger context window, and 5 match it exactly. Context is a capability limit rather than a cost one: paying more buys no extra room unless the larger window is what you are paying for.

Cached input reads at $0.006 against a standard input rate of $0.300, a 98% discount on the part of your prompt that repeats.

It accepts image input, but DeepSeek publishes no rule for how images become tokens, so we will not price a scanned page on it. That is a gap in the provider's documentation rather than a limit of the model, and it is the reason this model is excluded from image estimates in our estimator instead of being given a plausible figure.

Its 1M token context window holds roughly 750,000 words of English text in a single request. A workload that does not fit cannot run here at any price, which is a capability limit rather than a cost one.

Priced near this one

Closest to DeepSeek Flash on a retrieval workload, cheaper and more expensive alike. Cost only; we publish no quality score.

Retrieval profile: 800k in, 200k out, 70% cached where offered.