DeepSeek Adds Vision to Experimental V4 Flash Model
DeepSeek’s new API model can interpret images and screenshots for agentic tasks, but its performance claims remain based on the company’s own tests.
What you need to know
- DeepSeek-V4-Flash-Vision-Exp adds image and screenshot input to the text-only V4-Flash family.
- The experimental model is available via DeepSeek’s API Platform, rather than as a confirmed consumer app feature.
- DeepSeek says it approaches Claude Opus 4.8 on multimodal agents, but its own table shows it trailing on eight of 11 tests.
DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental version of its V4 Flash model that can interpret images and screenshots alongside text. Announced on Friday 21 August, the model is available through the DeepSeek API Platform and is aimed at software agents that need to work with visual information as well as written prompts.
The new release is not a confirmed feature for the standard DeepSeek website or mobile app. Instead, it is an API product for developers and businesses building their own tools. DeepSeek has also updated its agent environment, DeepSeek Harness 0.1.1, with support for the vision model.
The formal model identifier is deepseek-v4-flash-vision-exp. DeepSeek’s documentation lists support for text and image input, text output, tool calls, JSON output and the Responses API. It can be accessed through OpenAI-compatible and Anthropic-compatible API interfaces.
V4 Flash gains image input
Earlier DeepSeek-V4-Flash releases were text-only. The new vision variant can bring screenshots and other images into the same workflow as reasoning and tool use, rather than requiring a separate image-analysis service before an agent can act on the result.
DeepSeek says the model matches V4-Flash’s text abilities, including “agents, reasoning, and world knowledge”. It further claims the release is “a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8”. Those statements, and the benchmark figures accompanying them, are DeepSeek’s own claims rather than independent evaluations.
“This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities, including agents, reasoning, and world knowledge.”
The model has a one-million-token context length and a maximum output of 384,000 tokens, according to DeepSeek’s API documentation. It offers both non-thinking and thinking modes. FIM, or fill-in-the-middle, completion is not supported. The listed concurrency limit is 2,500 requests per account, subject to DeepSeek’s wider rate-limit rules.
How its Claude comparison stacks up
DeepSeek chose Anthropic’s Claude Opus 4.8 as its headline comparison. In the company’s 11-test table, V4-Flash-Vision-Exp beat Opus 4.8 on three benchmarks: DeepSWE by 1.3 points, Agents’ Last Exam by 1.6 points and ZeroBench by one point.
It trailed Opus 4.8 on the other eight tests. The largest stated gap was on NL2Repo, where DeepSeek’s model scored 57.7 against 69.7, a 12-point deficit. It also scored 63.6 to Opus 4.8’s 71.7 on DSBench-Hard, while the Terminal-Bench 2.1 result was closer at 83.9 versus 85.0.
DeepSeek’s comparison with the text-only V4-Flash shows the vision model ahead on six of seven text evaluations. That needs some context: the company notes that the older model ignored multimodal parts of ApexBench and Agents’ Last Exam because it could not process images. Giving the new model visual input therefore accounts for part of that apparent improvement.
The table does not include Anthropic’s Claude Opus 5, which was released before this announcement. Performance against that model has not been confirmed by the sources available.
Pricing, image handling and UK availability
DeepSeek lists the new model at the same token rates as V4-Flash. At off-peak times, cached input costs $0.007 per million tokens, uncached input costs $0.22 per million tokens, and output costs $0.66 per million tokens. The respective peak rates are $0.014, $0.44 and $1.32 per million tokens.
Peak hours are listed as 01:00-04:00 and 06:00-10:00 UTC. From 00:00 Beijing time on Sunday 23 August, DeepSeek says off-peak pricing will apply throughout Saturdays and Sundays under its revised weekend billing rule.
Images are converted into input tokens and billed with the text in a request. DeepSeek says an image can consume up to 384 tokens. Developers can send images as base64 data, external URLs or through the Files API, which allows an image to be uploaded once and reused in later requests.
No official UK-pound pricing is listed. The eventual amount paid by a UK customer could depend on currency conversion, VAT and any charges imposed by a payment method or intermediary, none of which are confirmed in DeepSeek’s published pricing.
What happens next
For ordinary buyers, the immediate change is limited: this is not a new consumer hardware product or a confirmed app upgrade. The more noticeable effect may arrive through third-party services that use DeepSeek’s API to read screenshots, process documents or help agents navigate visual software interfaces.
DeepSeek’s low listed token prices could make those uses attractive to developers, particularly where an image only forms part of a larger text-and-tool workflow. But the model remains explicitly experimental, and its published results do not show across-the-board parity with Claude Opus 4.8. Any company considering it for important work should weigh its own testing more heavily than a single vendor benchmark table.
Why it matters
This is chiefly a developer release, so UK users should not expect a new vision button to appear in the standard DeepSeek app or website. It could, however, make image-capable tools such as screenshot assistants and document-processing services cheaper to build, with images charged as input tokens rather than through a separate image fee. The experimental status and DeepSeek’s mixed self-reported benchmark results mean businesses should test it for their own workflows rather than treating it as a proven Claude replacement.