DeepSeek caps images at 384 tokens: reprice your OCR

DeepSeek shipped its first multimodal model on August 21: deepseek-v4-flash-vision-exp. The interesting number is not a benchmark. It is the billing rule. Images go in as ordinary input tokens at the same rate as text, with a hard ceiling of 384 tokens per image. No vision surcharge, no per-image fee. If you run a document pipeline — invoices, receipts, packing slips, ID scans — the DeepSeek vision pricing changes what that pipeline is allowed to cost.

What actually happened

The vision model is an experimental variant of V4-Flash-0731 that adds image understanding while holding the base model's text behavior. It shares V4-Flash's rate card exactly, per DeepSeek's pricing page: $0.007 cache-hit input, $0.22 cache-miss input, and $0.66 output per million tokens off-peak, doubling to $0.014 / $0.44 / $1.32 during peak hours (01:00–04:00 and 06:00–10:00 UTC).

The image handling is documented plainly. Images are resized before inference — small ones scaled up toward 384×384, large ones scaled down toward roughly 800×800 — and the vision guide states an upper bound of 384 tokens per image regardless of the original size. A 5000×5000 scan and a 400×400 thumbnail bill the same. Limits are generous: up to 600 images per request, 32 MiB per image via base64 or URL, 64 MiB via the Files API, JPEG/PNG/GIF/WebP, with format detected from file content rather than the filename.

Run the arithmetic. A thousand images is 384,000 input tokens at most. Off-peak cache-miss, that is about $0.08. At peak rates, about $0.17. Output tokens are separate and, for structured extraction, usually small.

Why cheap vision tokens matter for your business

Document work stops being a project and becomes a cron job. The reason most small operators never automated invoice intake was not capability — it was that per-page OCR pricing made a back-catalog run cost more than the labor it replaced. At a tenth of a cent per page, reprocessing five years of scans is a lunch break.

The 384-token ceiling makes cost predictable, which matters more than cheap. A flat cap per image means your bill scales with document count, not with whatever resolution your scanner happened to be set to. You can quote a customer a per-document price and be right.

It also caps how much detail the model gets. A dense multi-column invoice compressed to 384 tokens is not the same read as a full-resolution pass. Crop before you send. One region per call — line items here, totals there — beats one call on a busy page. Test against your worst documents, not your cleanest.

Batch off-peak and the rate halves. Peak is 01:00–04:00 and 06:00–10:00 UTC. Nightly document runs from Phoenix land squarely off-peak by default. If yours do not, move them.

Experimental means experimental. DeepSeek has already retired model aliases once, and the -exp suffix is a warning label. Put this behind your own abstraction — a function that takes an image and returns structured JSON — so swapping the model is a config change, not a rewrite. Keep the raw images. The extraction is reproducible; the source document is not.

Sources: DeepSeek API pricing, DeepSeek vision guide, OpenRouter: DeepSeek V4 Flash Vision Exp.