DeepSeek adds vision to V4-Flash: 1M-context image input, live now
The experimental model matches Opus 4.8 on some multimodal benchmarks — and OpenAI now lets eligible API keys pick a processing region per request.
Copy markdown
Vision at V4-Flash's price
Call model='deepseek-v4-flash-vision-exp' to feed images into DeepSeek's V4-Flash for document, chart, and multimodal-agent work — 1M context, images billed at the base model's token rate (capped 384 tokens/image, ~$0.22/M input on OpenRouter). By DeepSeek's own eval it edges Opus 4.8 on Agents' Last Exam (27.3 vs 25.7) and ZeroBench (35.0 vs 34.0); it's API-only, not open weights.
Pick a processing region per request
Eligible OpenAI project keys can now route individual requests to a specific processing region via prefixed API domains, instead of pinning a whole account or project. Handy when only some calls carry data-residency constraints but you don't want to split projects to satisfy them.