DeepSeek's Vision API Lands at 0.12 Cents Per Image — but 800×800 Cap and CDN Blocks Bite
At 1.15 yuan per thousand images, screenshot-heavy automation — UI testing, agent step logging, page monitoring — flips from a cost-prohibitive luxury to a default. The 800×800 ceiling and CDN restrictions mean the first generation is for layout-level understanding, not fine OCR, but the price alone resets what's practical to build.
The `deepseek-v4-flash-vision-exp` model quietly appeared in API docs, giving DeepSeek's API the ability to see for the first time. Pricing matches the text-only V4 Flash tier: images are scaled to roughly 800×800 and capped at 384 tokens, which works out to about 0.12 cents per image at peak rates. That undercuts Claude Sonnet 4.6 by a factor of 25 on per-image cost, making high-frequency screenshot analysis economically viable for the first time.
Early testing reveals sharp edges. Text recognition handles main content but mangles symbols, emoji, and currency signs. The model miscounts fingers in a hand sketch — a known multimodal failure mode. More critically, the `reasoning_effort` parameter appears non-functional, so thinking-mode token overhead cannot be turned off. Images hosted on Alibaba CDN (alicdn) and ByteDance CDN (bytetos/bytescm) fail with download errors due to hotlink protection, forcing a local-download-then-base64 workaround.
A Files API launched alongside the vision model, free for now, letting callers upload images once and reference them by ID across requests — useful for Agent workflows that revisit the same design mockups repeatedly.
DeepSeek's pricing strategy treats vision as a commodity input rather than a premium feature — no multimodal surcharge, same token rate as text — which pressures competitors whose vision tiers carry steep markups.
The 800×800 resolution cap is a deliberate trade-off that keeps per-image token counts low and costs predictable, but it also draws a hard line: this model is for layout and trend understanding, not for reading fine print or dense tables.
CDN hotlink failures reveal a practical gap between API design and China's web infrastructure; any service that relies on Alibaba or ByteDance CDNs for image hosting will hit this wall and need a preprocessing step.
The non-functional `reasoning_effort` parameter suggests the vision model was shipped before full integration with DeepSeek's thinking-mode controls, which means early adopters pay a token tax they cannot opt out of.