DeepSeek V4-Flash-Vision-Exp Adds Vision to V4-Flash—But Its Opus 4.8 Claims Need Testing

The result should be read narrowly. “Close to Opus 4.8” describes performance in DeepSeek’s selected evaluation set; it does not establish equal quality across every visual task, coding workload, or long-running agent session. The official changelog provides several Vision-Exp scores, including 36.5 on ApexBench, 27.3 on Agents’ Last Exam, 64.3 on Chartography, and 35.0 on ZeroBench Pass@5.

These figures also do not answer important production questions. Benchmark outcomes can change with prompts, tool environments, task selection, latency requirements, failure handling, and evaluation methodology. There is not enough independent evidence in the available material to treat the comparison as a definitive model ranking.

The text benchmark story is more conservative

The strongest claim for text-heavy use is continuity rather than clear dominance. DeepSeek positions Vision-Exp as retaining the existing V4-Flash model’s text capabilities, so developers do not have to give up its established reasoning and agent behavior simply to gain image input.

That distinction is useful when choosing a model:

• For text-only applications, Vision-Exp is presented as an extension of V4-Flash rather than a proven replacement for higher-end models.
• For visual agents, the added image understanding is the main reason to test it.
• For production decisions, the published scores should be supplemented with task-specific evaluations using the same tools, prompts, and latency targets as the intended application.

API access and image input

DeepSeek makes Vision-Exp available through its API platform. The release supports Chat Completions, Messages, and Responses, including mixed text-and-image input.

Images can be provided in three ways:

• as a base64 data URL;
• as a publicly accessible HTTP(S) URL; or
• as a file_id for an image uploaded through the Files API.

The documented image formats are JPEG, PNG, GIF, and WebP. The Responses API likewise accepts an input_image content part with an image URL, base64 data URL, or uploaded file reference.

Pricing and the 384-token image cap

Image input is converted into billable tokens. DeepSeek says each image contributes no more than 384 input tokens, using the V4-Flash pricing structure.

Published V4-Flash rates associated with the launch are:

• Uncached input: $0.22 per million tokens off-peak and $0.44 at peak
• Cached input: $0.007 per million tokens off-peak and $0.014 at peak
• Output: $0.66 per million tokens off-peak and $1.32 at peak

The 384-token ceiling makes the image portion of a request relatively predictable. It does not cap the cost of the accompanying text, tool calls, or generated output, so a complete agent interaction can still use substantially more tokens than the image itself.

Harness and Files API support

DeepSeek released Harness 0.1.1 with out-of-the-box support for Vision-Exp. That is relevant for developers building agents rather than issuing isolated image questions, because the model is designed to operate inside tool-using workflows.

The launch also introduced a free Files API. Developers can upload an image once and reuse it in later requests by passing a file_id, rather than repeatedly sending the same image bytes. DeepSeek documents file references in the file-api-... format.

Reusable file references are particularly useful when an agent needs to inspect the same screenshot, document page, or visual asset across multiple turns. The Files API itself may be free, but model inference and token usage remain billable.

What the launch means for AI competition

Vision-Exp shows how competition is shifting from model quality alone toward the full developer package: multimodal capability, API compatibility, pricing, reusable files, and agent tooling. DeepSeek is offering a visual extension of its Flash line at the same published token rates as V4-Flash, while Anthropic’s Opus 4.8 serves as the high-end comparison point in DeepSeek’s reporting.

The model’s practical importance will depend less on a single benchmark table than on whether developers find it reliable for screenshot-based coding, document analysis, interface agents, and visual tool use. Its experimental label is a reminder to validate those workflows directly before using it for critical production tasks.

The clearest conclusion is therefore measured: DeepSeek has made low-cost visual reasoning easier to access through its existing API ecosystem, and its early benchmark claims suggest a meaningful challenge in selected multimodal tests. Independent evaluations, real-world reliability, and sustained developer adoption will determine whether Vision-Exp becomes a durable alternative or remains an intriguing experimental release.