Skip to main content
2026/8/21 23:41 (UTC+8) · New Model · DeepSeek 🚀 DeepSeek’s first vision model deepseek-v4-flash-vision-exp is live, at no vision premium Built on the V4 Flash base with image input added, it keeps the 1M context, thinking mode, function calling and context caching, and is priced exactly like the text-only version: $0.44 input, $1.32 output per 1M tokens. Images convert to input tokens by size, capped at 384 per image, with large images rescaled to roughly an 800×800 equivalent —— 2000² and 4000² cost exactly the same, so pre-compressing saves bandwidth but not money. Images can be sent inline as base64 or by public URL (32 MiB per image). Both protocols work, and the group must match the protocol:
  • OpenAI format (/v1/chat/completions, /v1/responses): use a default group token
  • Anthropic format (/v1/messages, incl. Claude Code and similar clients): use a ClaudeCode group token
Both groups are priced identically —— the group affects capability only, never billing. We ran 124 test cases across roughly 1,100 calls; the three image channels, four formats and the detail token savings are all written up in the docs. 📖 DeepSeek V4 Flash Vision overview
← Back to Live Updates · 📚 Monthly archive