DeepSeek drops experimental vision model rivaling Opus

Adding vision to AI agents has always been notoriously expensive. DeepSeek is trying to flip that script with extreme efficiency.

DeepSeek launched an experimental multimodal model called DeepSeek-V4-Flash-Vision-Exp on its API platform. Developers can already call it using deepseek-v4-flash-vision-exp, and the newly updated DeepSeek Harness 0.1.1 supports it out of the box.

Vision agent power on a budget

On text tasks, the new model matches DeepSeek-V4-Flash across reasoning, agent tasks, and world knowledge without missing a beat.

The real upgrade is visual agency. On multimodal agent benchmarks, DeepSeek claims the model takes a massive leap over the text-only V4-Flash, coming remarkably close to Opus-4.8 performance.

The pricing design is just as notable. It bills images at V4-Flash rates with a hard cap of 384 tokens per image. Paired with a newly launched Files API, users can upload an image once and reference it repeatedly via file_id instead of re-uploading and repurchasing each time.

Why it matters

High-frequency multimodal agents that need to inspect UI screenshots every few seconds usually burn through developer budgets fast. DeepSeek capping image cost at 384 tokens while allowing file reuse looks like a direct fix for this bottleneck.

While this is still an experimental Exp release and its benchmark gains are specific to agent tasks, we think this sets a tough benchmark for competitors on cost per visual interaction.