DeepSeek-V4-Flash-Vision-Exp: vision at no premium

DeepSeek published DeepSeek-V4-Flash-Vision-Exp on August 21, 2026. benchr rechecked the pricing table and the changelog on September 7. The interesting part is not that a vision model exists; it is that adding vision cost nothing on the rate card, and that the thing you give up in exchange is not money.

Vision at no premium is unusual, and here it is exact

Most providers charge for image understanding, either through a separate model or a higher token rate. DeepSeek does neither. Every published cell matches DeepSeek-V4-Flash.

The second figure in each cell is the off-peak rate. DeepSeek halves every price during 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday — a five-hour window that lands in the European working morning and the small hours in the Americas. A batch job that can wait for it pays $0.22 to read and $0.66 to write.

Images are tokenized by dimension and billed as input tokens, so image cost is a function of resolution rather than a separate line. Large screenshots are expensive in the same way long documents are expensive; downscaling before the call is a real cost lever.

The word "experimental" is doing the work

DeepSeek calls the checkpoint experimental in its own changelog. That is not a disclaimer to skim past. An experimental endpoint carries no deprecation commitment, which means it can change behavior or disappear without the notice period a stable model would get.

benchr also found no license and no open-weights release for this checkpoint on any official page. DeepSeek-V4-Flash ships weights under MIT; the vision checkpoint does not, or at least does not say so anywhere benchr could read. The record carries null rather than inheriting MIT from its sibling, because inheriting a license is exactly the kind of assumption that turns into a compliance problem later.

Both facts point the same way: this is a model to prototype against, not one to build a contract around.

What DeepSeek claims it does better, and what it does not claim

DeepSeek's own framing is precise, and worth reading as written. On pure-text capability — agent, reasoning, world knowledge — it says the vision checkpoint is on par with the official DeepSeek-V4-Flash. On agent benchmarks requiring visual understanding, it says the checkpoint delivers a significant leap over V4-Flash.

So the claim is not a better model. The claim is the same model, plus eyes. If your workload has no images in it, DeepSeek is telling you there is no reason to move.

The published numbers, and whose they are

Nine figures come with the release. All nine are DeepSeek's own, measured by DeepSeek; benchr has not reproduced any of them.

The spread is the useful part. Terminal-Bench at 83.9 and AutomationBench at 25.7 are both agent benchmarks, and the gap between them says more about how differently those two suites are constructed than about the model. Read any single number here as one suite's opinion, not as a capability rating.

Who should try it, and who should not

If you do pilot it

Cost the images, not the calls. Because tokens scale with dimension, the same pipeline can cost several times more on full-resolution screenshots than on downscaled ones, and that ratio is set by your capture step rather than by DeepSeek. Measure tokens per image at the resolution you actually intend to send.

Then decide whether the off-peak window is reachable. A five-hour weekday discount is worth building a queue for if the work is batchable, and worth ignoring entirely if it is interactive. Halving the rate changes the economics of a document backlog; it changes nothing for a chat feature.

And pin the model string. deepseek-v4-flash-vision-exp is an experimental identifier, so an alias that silently moves is exactly the risk here — DeepSeek has retired aliases before, which benchr has covered separately.