DeepSeek gave Flash eyes: up to 384 tokens per image at the same price — but it is still experimental

The bottom line

DeepSeek has finally put native image understanding into its low-cost Flash API, and the most disruptive number is not a benchmark score: each image consumes no more than 384 tokens.

On August 21, DeepSeek-V4-Flash-Vision-Exp went live on the official API. It accepts mixed text-and-image input directly instead of requiring another model to describe the image first. It uses the same pricing as V4 Flash and works through Chat Completions, Anthropic Messages and the Responses API.

That gives screenshot analysis, browser operation, chart understanding and visual coding a strikingly inexpensive new entry point. The Exp suffix still matters, however. This is an experimental model, and DeepSeek's claim that its multimodal agent capability approaches Opus 4.8 is based on vendor-reported results. There is not yet enough independent testing to show that it is equally reliable in real agent workflows.

DeepSeek's reversal: Flash can finally see images directly

V4 Flash was previously a text model. Developers who wanted it to understand screenshots usually placed a separate vision model in front of it: the vision model converted an image into a written description, and Flash reasoned over that description. The workaround functioned, but it added latency, cost and information loss.

Now a developer can set the model to deepseek-v4-flash-vision-exp and send images directly. DeepSeek documents three input routes: inline base64 data, a publicly reachable image URL, or a file uploaded once and referenced later through the Files API.

The important target is agents, not merely image chat. A browser agent can inspect a page screenshot before deciding what to click. A coding agent can read an error screenshot or design mockup. An office agent can analyze charts, scans and presentation slides. Vision can now remain inside the same conversation and tool-call chain instead of living in a separate preprocessing service.

The real weapon is 384 tokens, not the word "multimodal"

DeepSeek resizes each image while preserving its aspect ratio and then converts the visual input into tokens. The official rule is unusually clear: one image consumes at most 384 tokens, so the charge does not keep rising indefinitely with the original pixel count.

The pricing table gives Vision-Exp the same rates as V4 Flash. One million uncached input tokens cost RMB 1.5 off-peak or RMB 3 during peak hours; one million output tokens cost RMB 4.5 off-peak or RMB 9 at peak. Image tokens and text tokens are billed together.

At the 384-token ceiling, the pure input charge for one image is tiny. The larger bill is more likely to come from an agent taking repeated screenshots, observing the page after every action, retrying failed steps and then producing a long textual answer. Looking once is cheap; letting an agent look a hundred times may not be.

Three API families work, while Files API stops agents moving the same image repeatedly

This release is not limited to one OpenAI-style endpoint. DeepSeek's documentation confirms image support across three API families:

• Chat Completions accepts image_url or file content blocks;
• Anthropic Messages accepts base64, URLs or Files API references and requires a beta header when a stored file is referenced;
• Responses uses input_image and can also receive images in tool outputs.

The Files API matters because it enables reuse, not because it adds another upload button. A design mockup, scan or product photograph can be uploaded once and referenced by file_id in later requests instead of being transmitted in every turn. DeepSeek allows an expiry from one hour to 30 days, or permanent storage when no expiry is supplied, and accepts individual files up to 64 MiB.

That convenience raises a governance obligation. Files belong to the API key, and permanent storage is not always desirable. Enterprise teams should set explicit expiry periods, restrict what can be uploaded and confirm whether customer images or internal documents are permitted to enter a third-party API.

Close to Opus 4.8 in benchmarks? First check who ran them

DeepSeek's changelog reports 83.9 on Terminal Bench 2.1, 59.3 on DeepSWE, 36.5 Pass@1 on ApexBench and 27.3 on Agents' Last Exam for Vision-Exp. The company says the model makes a large jump over text-only Flash on vision-dependent agent benchmarks and approaches Opus 4.8.

Those numbers are informative, but they do not mean that a 384-token DeepSeek image input has already become equivalent to Opus. First, DeepSeek reported the results itself. Second, the text-based code-agent tasks used DeepSeek Harness in minimal mode with max reasoning effort, temperature=1.0 and top_p=0.95. Third, the text-only Flash baseline ignores the multimodal elements in ApexBench and Agents' Last Exam, so a large gap between the old and new models is expected.

What can be confirmed today is that Vision-Exp closes Flash's most obvious capability gap. What cannot yet be confirmed is whether it matches the stability of leading proprietary models on dense OCR, tiny interface text, complex spatial relationships, long browser trajectories and hallucination control.

Our judgement: a low-cost ticket into multimodal agents

DeepSeek's smartest move was not building another model marketed as a flagship vision system. It attached vision to Flash, a model that is already inexpensive, offers high concurrency and is gaining a broader agent ecosystem.

Migration costs are also low. The existing DeepSeek base URL remains unchanged. Most integrations need to change the model name, represent message content as mixed text-and-image blocks and manage uploaded-file lifecycles. This is closer to unlocking a new API capability than rebuilding an application.

Our conclusion is cautiously positive: it is worth running a small A/B test on real work, but one vendor benchmark chart is not a reason to replace a production vision model immediately.

Who should test it now

• Teams already using V4 Flash that want agents to read screenshots or design mockups directly;
• Products that need to process product images, receipts, charts or browser screenshots at scale while keeping per-request cost low;
• Developers comparing agent integrations across Responses, Messages and Chat Completions.

What to watch

• Whether independent evaluations reproduce the reported 83.9, 36.5 and 27.3 results;
• How much the 384-token visual compression loses on tiny text, dense tables and high-resolution interfaces;
• Version stability, rate limits, latency and whether the experimental model becomes a production release;
• Whether the Files API's free-upload policy, storage governance and long-term retention rules change.