DeepSeek Ships Its First Vision Model, Matching Opus-4.8 on Agent Benchmarks at Flash Pricing
Agent developers who previously had to route images through a separate vision model before feeding text to DeepSeek can now run end-to-end visual reasoning in a single call, at the same price as the text-only model. The Opus-4.8-level benchmark scores and 384-token image cap make this a low-risk drop-in for screenshot analysis, UI automation, and document understanding pipelines.
DeepSeek's first multimodal model, V4-Flash-Vision-Exp, launched as an experimental release that keeps V4-Flash text performance intact while adding vision. Official benchmarks show multimodal agent scores approaching Opus-4.8, closing a gap that previously forced developers to chain separate vision models into DeepSeek workflows. The model handles JPEG, PNG, GIF, and WebP inputs through an OpenAI-compatible API, with images capped at 384 tokens regardless of resolution.
Pricing matches V4-Flash exactly, with off-peak input cache hits at 0.05 CNY per million tokens. A free Files API lets callers upload images once and reference them across requests, and the model also works through Anthropic and Responses API endpoints. DeepSeek Harness and Codex integrations are already updated.
Three practical prompt templates ship alongside the release: UI-screenshot-to-Tailwind-code, chart data extraction with trend analysis, and whiteboard architecture sketch to structured Markdown. The "Exp" label signals ongoing iteration, but the pricing and benchmark parity with Opus-4.8 make it immediately usable for agent pipelines that need end-to-end visual reasoning.
DeepSeek chose to ship vision as an additive module on top of an existing text model rather than as a separate multimodal model, which preserves the text performance developers already rely on and avoids the regression risk that often accompanies new modality support.
The 384-token image cap is aggressive — a 1920×1080 screenshot already hits 369 tokens — which keeps per-image costs predictable but means high-resolution detail work may hit the ceiling quickly.
Supporting Anthropic and Responses API endpoints alongside OpenAI compatibility signals that DeepSeek is targeting agent-framework portability, not just raw API consumption, making it easier to swap into existing Claude-based pipelines.
Labeling the model 'Exp' while publishing benchmarks against Opus-4.8 is a soft launch tactic: it sets expectations for iteration while demonstrating production-grade capability, lowering the barrier for developers to start building on it now rather than waiting for a GA release.