DeepSeek just shipped its first multimodal API model — DeepSeek-V4-Flash-Vision-Exp — and priced it like a loss leader: one image counts as at most 384 tokens, billed at the same rate as text. That works out to roughly 1.15 yuan (~$0.16) for 1,000 images at peak pricing, versus about $4 on Claude Sonnet 4.6 — a 25x gap that quietly turns "should this agent stop and look at the screen?" into a non-question.
What DeepSeek Just Shipped
DeepSeek-V4-Flash-Vision-Exp is available in the API immediately and is billed identically to the text-only V4-Flash: 3 yuan per million input tokens, halved again during off-peak hours. The vision capability is a cap, not a surcharge — a single image can consume at most 384 tokens, no matter how dense it is.
The model is also stronger on text than the version it extends: across seven agent benchmarks it beats V4-Flash-0731 on six of them, and on some multimodal tasks it edges out the top Western models. In plain terms, you no longer need a second, separate model just to make an agent read images.
Hands-On: From Screenshot to Shippable Code
In a live test, the model was handed a screenshot of the OpenAI homepage and asked to reproduce it as a runnable single-file HTML page — describe the layout, fonts, hex colors, and spacing rhythm first, then write the code, copy every word verbatim, restore hover states, scroll fades, and button textures, and stay intact at 375px. The replica came back near-identical.
A second, deliberately different test asked for an 11-page B2B deck for a specialty coffee roastery — light-roast single-origin beans only, a 5 kg minimum order, three real supply tiers, and a craft, slightly stubborn tone with zero third-wave-coffee jargon. The model delivered a client-ready proposal: specific numbers, no industry buzzwords, nothing written for outsiders. A vague brief in, a deliverable ready to send out.
The Price Math Behind "1 Yuan for 1,000 Images"
Vision providers bill per image token, and the spread is enormous. Per TokenMix's vision API benchmark, reading one 1024×1024 image consumes:
- Claude Sonnet 4.6: 1,334 tokens — about $4 per 1,000 images (~29 yuan)
- GPT-5.4 Vision: 765 tokens — about $1.9 per 1,000 images
- Gemini 3.1 Pro: 258 tokens — about $0.5 per 1,000 images
- DeepSeek V4-Flash Vision-Exp: capped at 384 tokens
At V4-Flash pricing, 1,000 images cost about 1.15 yuan at peak and under 0.6 yuan off-peak. That is a 25x gap against the most expensive option — and closer to 50x during off-peak hours.
That number is the whole story: 29 yuan and 1 yuan for the same 1,000 images are two different product designs. The expensive version forces you to decide whether a step is worth looking at; the cheap one lets an agent screenshot every step and check its own work. (For a sense of how much vision quality can vary between models, our GPT-5.6 Sol vision benchmark is a useful cross-check.)
Completing the Agent Loop in Eight Days
DeepSeek has been assembling the full agent stack in a hurry. On August 13 it open-sourced DeepSeek Harness with native Responses API support — the runtime half. On August 21 vision went live across Chat Completions, Messages, and Responses — the input half — and Harness 0.1.1 shipped the same day with out-of-the-box support for the new model, so no adapter code is required.
Images can enter three ways: base64 inline, an external URL, or the newly opened Files API, which is free and lets you upload once then reference by file_id. For agents that re-read the same design file, report, or screenshot repeatedly, that removes the round-trip cost of every turn. If you are new to the runtime, our Harness plugin install guide walks through the ecosystem.
Why 25x Cheaper Multimodal Is a Structural Change
Multimodal input has always been priced as a premium add-on. DeepSeek just made it an included capability, and that shifts the cost assumptions underneath agent design. When vision is nearly free, "look at your own output" becomes a default verification step rather than an optimization — agents that can self-check their work visually behave differently from ones that only read text back.
It also removes a build headache: teams no longer route images to a second provider just for OCR or UI understanding. One endpoint, three API formats, three input paths — the friction that kept vision out of agent loops is gone.
What You Can Do With It
- Add visual self-checks to agents. Screenshot a rendered page and have the model verify layout, colors, and content — it now costs cents at scale.
- Use the Files API for repeat reading. Upload a design or report once, reference by file_id, and stop re-encoding base64 every turn.
- Re-run your vision budget math. If you pay per-image surcharges elsewhere, benchmark your OCR and document workloads against this endpoint.
- Test the off-peak tier. At 50% off, batch jobs and nightly pipelines get meaningfully cheaper.
FAQ
Q: How much does DeepSeek V4-Flash Vision cost per image?
A: One 1024×1024 image counts as at most 384 tokens, billed at V4-Flash rates (3 yuan per million input tokens). That is about 1.15 yuan for 1,000 images at peak pricing and under 0.6 yuan off-peak — roughly 25x cheaper than Claude Sonnet 4.6.
Q: Which API formats does DeepSeek-V4-Flash-Vision-Exp support?
A: It supports Chat Completions, Messages, and Responses API formats, with three image input paths: base64 inline, an external URL, and the new free Files API.
Q: When did DeepSeek-V4-Flash-Vision-Exp launch and is it available now?
A: It launched on August 21, 2026 and is available immediately in the API. DeepSeek Harness 0.1.1, released the same day, supports it out of the box with no adapter code.
