On August 21, DeepSeek released deepseek-v4-flash-vision-exp on its API platform; developers can call it by simply setting the model parameter to that name. It's the first vision-capable model in the V4 lineup, and DeepSeek has tagged it Exp, explicitly flagging it as an experimental preview that isn't recommended for production deployments yet.
On the text side, capabilities line up with the standard V4-Flash release. According to DeepSeek's official update page, the model scores 83.9 on Terminal Bench 2.1, 57.7 on NL2Repo, 59.3 on DeepSWE, and 64.3 on Chartography. On vision-related agent benchmarks, DeepSeek says ApexBench, Agents' Last Exam, and Chartography all show clear gains, with multimodal agent performance approaching Opus-4.8. The choice of benchmark target is telling — Anthropic's flagship has long been the yardstick for screen-operation tasks.
One image, ¥0.000384
Pricing is the more interesting part of this release. The vision model keeps V4-Flash's existing price sheet: ¥1 per million input tokens, as low as ¥0.02 per million on cache hits, and ¥2 per million output tokens. Images aren't billed separately — they're converted into tokens and counted as input, with each image capped at 384 tokens.
Do the rough math and the vision-input cost of a single image comes out to roughly ¥0.000384, meaning about 2,600 images cost a single yuan. GUI agents running a moderately complex task routinely take dozens to hundreds of consecutive screenshots; at 100 screenshots, the image-related spend works out to roughly ¥0.038. Folded into a full task bill, that's practically negligible — the real cost center shifts back to text reasoning and output length.
Per-image pricing for vision APIs takes a different route elsewhere in the industry. Charging per image is predictable, but for agent-style applications that hammer screenshots as a high-frequency sensor input, the bill scales linearly with call volume. Converting images to tokens and capping that count instead pins the cost of a single frame to a very small, fixed number.
The flip side of the cap
384 tokens isn't a high number. Before an image enters the model, it gets sliced into patches and encoded — the token budget directly determines how finely the model can see. Capped at 384, a large image gets compressed into a fairly coarse representation. That's probably enough to read the overall layout of a webpage screenshot, but reading small type in a dense table, or fine details like a status bar on a phone screenshot, is far less certain to hold up.
DeepSeek itself hasn't oversold this either, saying only that it will keep tuning stability and output quality based on real-world feedback. For developers, the advice is to first throw the messiest images from your own use case at it — cluttered admin dashboards packed with small text, scanned documents, watermarked charts — rather than judging it purely by benchmark scores. The context window remains 1 million tokens, and thinking effort still comes in two tiers, high and max.
The last piece of a four-month puzzle
Laid out on a timeline, DeepSeek's pace this year comes into focus. V4 launched on April 24, putting the old deepseek-chat and deepseek-reasoner names on a deprecation countdown; V4-Flash shipped separately on July 31; V4-Pro added native Responses API support and Codex compatibility on August 13, split thinking effort into three tiers — low, high, and max — and rolled out peak/off-peak pricing starting August 17; and now, on August 21, the vision piece has been filled in.
Four iterations in four months is a tight cadence. None of them centered on parameter count or topping leaderboards — the focus each time has been on calling conventions and billing structure: the Responses API, tiered thinking effort, peak/off-peak pricing, and now a per-image token cap. All of them are dials aimed squarely at developers' cost calculations. Shipping the vision model with an Exp suffix first follows the same playbook — get it into people's hands, collect feedback, and talk about a stable release later.
Sources: DeepSeek API official update page, CocoLoop, Kuaikeji; benchmark scores and pricing follow DeepSeek's official documentation, with the per-image token cap and input/output/cache pricing cross-checked against both sources.