DeepSeek added vision to its cheapest model and charged nothing extra for it. The company's API documentation confirms that deepseek-v4-flash-vision-exp is live, accepts images as inline base64 data, as a URL, or as a file uploaded through the Files API, and is billed at standard V4-Flash pricing. Image understanding at DeepSeek's price point is a meaningful change in what the cheap tier of the market can do.
Key facts
-
deepseek-v4-flash-vision-expis live, per DeepSeek's API release notes. - Images can be supplied as base64, URL, or Files API upload.
- It is billed at V4-Flash pricing, identical to the text-only model.
- The
expsuffix marks it experimental; the vision guide documents the interface.
Vision used to be a premium feature. When multimodal models first appeared, image understanding was a flagship capability, priced accordingly and available only at the top of a provider's lineup. The progression since has been steady and one-directional: from flagship-only, to available across the lineup, to available on the budget tier, to available on the budget tier at no premium. DeepSeek's release is the last step of that sequence for one of the cheapest capable models in wide use.
The three input methods sound like a footnote and are not. Base64 means you can embed an image directly in a request without hosting it anywhere, which is what you want for a desktop application or a script processing local files. A URL means you can point at an image the model fetches itself, which is what you want when the images already live in object storage and you would rather not pull gigabytes through your own service to push them back out. The Files API means you can upload once and reference many times, which is what you want when the same document gets asked about repeatedly. Each covers a genuinely different integration shape, and providers that support only one of them force awkward workarounds.
What makes cheap vision interesting is the class of work it opens up. Reading a screenshot, extracting a table from a scanned invoice, checking whether a photo shows what a form claims it shows, describing a chart in a report: these are all high-volume, low-value-per-item tasks. At flagship pricing they do not pencil out, because the value of correctly reading one invoice is less than the cost of the call. At commodity pricing they do, and that shift is where most of the practical deployment of vision models is going to come from. It is not the impressive demos. It is the boring pipeline that used to require optical character recognition software and a lot of glue.
The reason vision is affordable to add at all comes down to how these models process images. A picture is converted into a sequence of tokens, much as text is, and then flows through the same layers as everything else. There is no separate image model running alongside the language model in current designs; the vision component is a comparatively small encoder feeding into the same stack. That is why a provider can offer image input without a price change. The marginal cost is the tokens the image consumes, and those are billed like any other input tokens.
Why this matters in context: DeepSeek has spent two years being the company that makes the expensive thing cheap, and the pattern here is the same one. The frontier labs establish a capability, and then somebody demonstrates that the capability does not require frontier pricing. That compresses margins across the market, which is uncomfortable for vendors and excellent for anyone building on top of them.
Two honest caveats. The first is the exp in the model name, which is not decoration. An experimental endpoint can change its interface, change its behaviour, or disappear, and anything built on it should be built with that in mind. The second is that the release notes and the vision guide document what the model accepts, not how well it performs. There is no published accuracy figure here, no benchmark comparison against the multimodal models it undercuts, and no statement about how it handles the failure cases that plague vision models generally, such as dense text in low-resolution images, unusual chart types, or images that contain instructions the model might follow. That last one is not a hypothetical: image-borne prompt injection is a live attack class, and a cheap vision model pointed at untrusted images inherits every bit of it.
So the accurate read is narrow and still useful. DeepSeek has made image input available at a price where high-volume use is economically sensible, on an experimental endpoint, without telling anyone how good it is. Whether it is good enough for a given pipeline is a question each user will have to answer by testing, which is how it usually goes at this end of the market.
Originally published on Ground Truth, where every claim is checked against the primary source.
