DeepSeek ships V4-Flash-Vision — multimodal at a cut price
DeepSeek's experimental vision model reportedly comes close to Opus 4.8 on agent benchmarks — one image costs at most 384 tokens.
Illustration · AI-generated (AI IN LIFE)
At a glance
- Release: August 21, 2026, model ID deepseek-v4-flash-vision-exp, live on the DeepSeek API
- Billing: at most 384 tokens per image at V4-Flash prices
- Company claim: multimodal agent performance "close to Opus 4.8"
- Chinese media: around 1,000 images from roughly 1 yuan
- Launched alongside: a free Files API and DeepSeek Harness 0.1.1
What's new? On August 21, 2026, DeepSeek switched on the experimental model deepseek-v4-flash-vision-exp on its API platform — the Chinese lab's entry into the multimodal agent market. According to the official release notes, the model combines V4-Flash's text strengths with new image understanding.
How good is it? DeepSeek's own wording: the model "matches DeepSeek-V4-Flash on text capabilities" — holding the text model's level on agents, reasoning and world knowledge — while bringing multimodal agent performance "close to Opus 4.8", Anthropic's current flagship. Independent benchmarks are still pending; Caixin and Seeking Alpha frame the claim as a direct challenge to the US labs.
What does it cost? Images are tokenised for billing: one image consumes at most 384 tokens at regular V4-Flash rates. Chinese media calculate that 1,000 images can be processed from roughly 1 yuan — a fraction of what Western frontier models charge.
What ships alongside it? A free Files API launched at the same time: upload an image once, get a file_id, and reference it across any number of requests without re-uploading. Chat Completions, Messages and Responses formats are supported, and DeepSeek's own tooling, DeepSeek Harness 0.1.1, arrives with built-in compatibility.
What does it mean for the market? DeepSeek is repeating its text-market price playbook in vision: near-frontier performance at radically low cost. For European companies piloting visual agent workflows — document or screenshot analysis, for instance — the entry barrier drops sharply, while the usual data-residency and compliance questions remain.
FAQ
Is the model production-ready?
It is flagged experimental (Exp); DeepSeek positions it for testing — independent benchmarks are still missing.
How does it differ from V4-Flash?
It adds image input and multimodal agent capabilities on top of the text model, at the same text performance and token prices.
How do images get into the model?
Via Base64, external URLs, or the new Files API with a reusable file_id.
