DeepSeek adds vision to V4-Flash: 1M context at $0.22/M input
Cheap enough for vision inside agent loops — plus a free Files API that reuses images by ID, while gpt-image-2 preview adds transparent PNGs.
Copy markdown
Vision in the agent loop, at V4-Flash prices
The experimental deepseek-v4-flash-vision-exp adds image input to V4-Flash 0731 — a 284B sparse MoE with 13B active params — at $0.22/M in and $0.66/M out over a 1M-token context, with images billed at up to 384 tokens each and no vision surcharge. DeepSeek's own charts put its multimodal-agent scores near Opus 4.8. It's live on the DeepSeek API and OpenRouter.
A free Files API stops re-uploading images
Shipped alongside the model, DeepSeek's Files API lets you upload an image once and reference it by ID across many requests at no extra cost — so a screenshot-heavy agent stops re-paying to resend the same pixels every turn. It works across the Chat Completions, Messages, and Responses endpoints.
gpt-image-2 can finally cut out the background
OpenAI's image model now accepts background: transparent in preview, returning real alpha-channel PNG or WebP instead of a baked-in backdrop. One parameter turns it into an asset pipeline — logos, stickers, UI icons, product cutouts — with no manual masking. Build this weekend: a batch generator that drops transparent art straight into your app.