DeepSeek launches vision model it says rivals Anthropic’s Opus 4.8

Chinese AI lab DeepSeek released an experimental multimodal model on Friday that it claims approaches the performance of Anthropic's Claude Opus 4.8 on key benchmarks, extending the company's push into visual AI at a fraction of the cost charged by American competitors.

A Vision Variant, Not a New Flagship

The model, called DeepSeek-V4-Flash-Vision-Exp, is a multimodal variant of the company's existing V4-Flash text model. According to DeepSeek, it retains the text capabilities of V4-Flash — including agentic behavior, reasoning, and general knowledge — while adding image understanding on top.officechai

On multimodal agent benchmarks published by the company, the new model narrows the gap to Opus 4.8 considerably. On ApexBench, V4-Flash-Vision-Exp scores 36.5 at Pass@1 compared to Opus 4.8's 39.4. On Agents' Last Exam, it edges past Anthropic's model with 27.3 versus 25.7, and on ZeroBench it scores 35.0 against 34.0. On text-heavy evaluations like Terminal Bench 2.1, the model trails Opus 4.8 by roughly one point (83.9 versus 85.0), though wider gaps remain on benchmarks like NL2Repo (57.7 versus 69.7).officechai

DeepSeek evaluated its own models using its internal Harness Minimal Mode, meaning the numbers have not been independently verified.officechai

Pricing and Developer Tools

The model is available now on the DeepSeek API under the identifier `deepseek-v4-flash-vision-exp`. Images are billed by token count, capped at 384 tokens per image, at the same rate as V4-Flash's existing pricing — a structure that keeps costs well below what U.S. labs charge for comparable capability.officechai

Alongside the model, DeepSeek shipped version 0.1.1 of its Harness framework with built-in support for the release, and a Files API that lets developers upload an image once and reference it repeatedly across requests using a file ID. The Files API is free to use.officechai

Competitive Context

The release comes as DeepSeek faces intensifying competition from both Chinese rivals like Alibaba ↗ 1.28% and Moonshot AI, and U.S. firms including Anthropic and OpenAI. Bloomberg reported Friday on the release, framing it as DeepSeek's latest effort to challenge Anthropic's position in advanced AI.bloomberg+1

Rather than a flagship launch, the experimental model fits a pattern among Chinese labs of extending multimodal and agentic capabilities into mid-tier models. Whether V4-Flash-Vision-Exp graduates from experimental status will depend on developer adoption — for now, it signals that the next round of AI competition is as much about making cheap models capable of visual reasoning as it is about training the largest systems.officechai