What DeepSeek charges now

Since 10 September, V4.1 Flash replaces V4 Flash. Prices are per million tokens, and DeepSeek has two time bands.

Peak is Monday to Friday, 01:00–04:00 and 06:00–10:00 UTC. In Italian summer time that lands on 03:00–06:00 and 08:00–12:00 — exactly the working morning. Everything else, weekends included, costs half.

Against the old V4 Flash ($0.22 input / $0.66 output off-peak), input drops 32% and output only 9%.

Why the saving is 17%, not 32%

We have a measured run to work from: a whole site optimised with One-Click SEO — 19 items in 1 minute 57, producing titles, meta descriptions, keyphrases, AI Context, image ALT texts and a Q&A set. Total: about 165,000 tokens in and 110,000 out.

The reason is simple arithmetic that price lists hide: input is only 27% of what you spend. Output costs four times as much per token, and SEO work is writing — the model reads a page once and then produces titles, descriptions and alt texts. The 32% cut applies to the smaller slice.

Which is still a saving worth having: optimising that site now costs about nine cents of a dollar. But we would rather give you the number that matches your invoice than the one that looks best in a headline.

Cache hit: $0.003 per million

When part of your input has been processed recently, it is billed at two hundredths of the normal input price. On the run above, that would take the total from $0.091 to $0.067 — another 27% off, without changing anything you do.

It matters more than it sounds, because AISA sends a stable instruction block before the content that changes: your Brand Context, your rules, your tone. That part repeats, so it is exactly what a cache is for.

Vision is now part of the same model

DeepSeek retired its experimental vision model. From V4.1 Flash, image understanding lives inside the text model: one model, one price, no separate setup.

We checked it against the live API rather than trusting the changelog, and generated a real ALT text through AISA. It works, and it is the cheapest image description in the catalogue. If you had selected the old experimental model, AISA now maps it to the current one on its own — nothing to change on your side.

Where this sits among the providers

AISA works with your own API keys, so the provider is your choice. These are the rates AISA uses to estimate your costs, per million tokens:

Off-peak, DeepSeek and Mistral are level — and Mistral is the only EU-based option, which matters to some. During peak hours DeepSeek doubles and slips behind ChatGPT. There is no single winner: there is the provider that fits when you work.

One caveat we would rather state than hide: a lower price per token does not always mean a lower bill. Different models count tokens differently for the same text, so the only figure that settles it is the one on your invoice.