DeepSeek priced image input at text rates: a close read of deepseek-v4-flash-vision-exp

DeepSeek released an experimental model called deepseek-v4-flash-vision-exp on August 21, 2026 that accepts image input while carrying exactly the same price table as the text-only deepseek-v4-flash. Per the official API documentation, input tokens cost $0.22 per million off-peak and $0.44 peak on a cache miss, output runs $0.66 off-peak and $1.32 peak, and a single image is billed at up to 384 tokens regardless of its resolution. A single request accepts up to 600 images with a maximum edge length of 8,192 pixels. ASAP summarizes the official DeepSeek API documentation as the primary source.

One image costs at most 384 tokens, and the billing unit came first

The first thing DeepSeek pinned down in the documentation is not performance but the billing unit. Images are normalized to roughly 800×800 pixels before tokenization, which caps a single image at 384 input tokens. Feeding in an 8,192-pixel image or an 800-pixel one hits the same ceiling.

The arithmetic that follows is simple. At the cache-miss off-peak rate of $0.22 per million input tokens, 1,000 images work out to about 384,000 tokens and roughly $0.09 in input cost. Running the same batch during peak hours at $0.44 brings it to about $0.17. DeepSeek defines peak hours as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, with every other hour billed off-peak at half the rate.

600 images, 8,192 pixels, and the rule that changes at 15

The per-request ceiling in the DeepSeek vision guide is 600 images at a maximum edge length of 8,192 pixels, with one condition attached: once a request contains 15 or more images, the maximum edge drops to 4,096 pixels.

Size limits also split by delivery method. Images placed inline in the request body cap out at 32 MiB, while a file uploaded through the Files API and referenced by ID can reach 64 MiB, and total image payload per request rises to 200 MiB when file IDs are used. For high-count workloads such as scanned documents or frame batches, the design steers you toward file upload rather than inline encoding.

The detail field buys tokens, not picture quality

Requests accept a detail field that changes how the image is processed. Setting it to low downscales the image to 512×512 for faster and cheaper inference, while high and original preserve full resolution, and the documentation states that auto currently behaves the same as original.

That last clause matters immediately in practice, because leaving the field unset and trusting the default routes the request down the most expensive path. In a large batch job, the difference between omitting detail and setting it to low is one line of code, but it is not one line on the invoice.

Matching the text model's price table is the real headline

What stands out in this release is not the specification but where the price table sits. Multimodal models usually land in a more expensive tier than the text-only model in the same family, yet deepseek-v4-flash-vision-exp matches deepseek-v4-flash on all three line items: cache hit, cache miss, and output. In the same document, deepseek-v4-pro lists output at $1.98 per million tokens off-peak, three times the flash tier.

The placement says image understanding has moved from a separate product into a baseline feature of the cheapest tier. Until now, image input was a capability teams evaluated on cost before capability, and workloads like document parsing or screenshot reading often ended with the conclusion that a person doing it manually was cheaper. A 384-token ceiling combined with flash rates changes the output of that calculation.

The cache-hit rate deserves equal attention. At $0.007 per million tokens off-peak, a cache hit is roughly one thirty-first the cost of a cache miss, so whether a pipeline re-references the same images or re-uploads them every time swings real spending by more than thirtyfold. The largest variable in the published table is cache design, not model choice.

A model shipped without benchmarks

One caution is unavoidable. DeepSeek's official API documentation carries no benchmark numbers for this model. Specifications, limits, and prices are precise, while any figure measuring capability is absent. The exp suffix in the model name and the description of an experimental model that additionally accepts image input are how DeepSeek positions it.

Reports circulated after launch claiming the model approaches top-tier systems on agent benchmarks, but ASAP could not verify those figures in DeepSeek's own documentation. Deciding on adoption from unverified performance claims is risky, and what is verifiable right now is the cost structure and the input limits, not the capability.

Before putting an exp model into production

An experimental label carries two properties at once: the model is cheap and openly accessible, and its behavior can change or disappear without notice. The documented note that auto currently matches original is itself an acknowledgment that this may change.

That points to where the model belongs. Low-cost-of-failure positions fit best, such as a first-pass filter for bulk document classification, screenshot-based log cleanup, or draft image metadata generation, where a person or another model reviews the output afterward. Work that reaches users directly or drives hard-to-reverse decisions is better deferred until benchmarks exist. Above all, because an exp model can be withdrawn at any time, it belongs behind a swappable adapter rather than inside code that depends on it directly.

Why a compatible API and a harness preview arrived together

In the same documentation, DeepSeek states that its API is compatible with the OpenAI and Anthropic formats and outlines integration with tools including Claude Code, GitHub Copilot, and OpenCode. Listed alongside is DeepSeek Harness, described as being in developer preview.

The combination is not incidental. Format compatibility lowers switching cost, and a harness is the execution layer that connects a model to real work. Recognition that the execution layer shapes outcomes as much as raw model capability has been surfacing across several camps at once, and this bundle points the same direction. Shipping cheap image input, an interface that drops into existing tooling, and a first-party execution layer together looks less like selling a model and more like claiming the slot the model will sit in.

What a team should calculate now

The workloads where this specification matters immediately are high-volume and judgment-light. Processing scanned contracts and invoices, checking shelf photography, reviewing ad creative, and running app screen regression tests all repeat across thousands of images, and per-image cost has been the deciding factor. The 600-image ceiling and the drop to a 4,096-pixel maximum edge at 15 or more images are the first numbers to check when sizing batches.

One category stays outside the documentation, though. Where the data goes, whether it is used for training, and whether images containing personal information may be sent are questions a price table does not answer. Scanned documents, store photos, and app screens all mix personal data with commercial secrets, and a low rate is not grounds for skipping that review. Cost calculation and data clearance are separate steps.

Open questions

Three questions about deepseek-v4-flash-vision-exp cannot be answered from what DeepSeek has published so far. The first is capability: how far an image representation compressed into 384 tokens can read tables, schematics, and small type has no official figure, which leaves direct measurement as the only option. The second is durability, since nothing in the documentation says whether an exp model graduates to a stable tier or is quietly retired.

The third is whether this price holds. Attaching image input at text rates may be an experimental-stage pricing decision, and a separate tier at general availability remains possible. A team evaluating the model now should finish capability and cost testing in the same week, then treat the result not as a verdict on this one model but as a measurement of what becomes possible once per-image input cost falls this far.

Source: DeepSeek API Documentation (api-docs.deepseek.com), models, pricing, and vision guide pages, summarized by ASAP.