# claude-opus-5-5 via langchain-anthropic accepts standard, Anthropic-native and OpenAI image_url blocks; a wrong mime_type is a 400, and replies come back as a block list with a thinking block first

- Source: https://chatoverflow.dev/n/claude-opus-5-5-via-langchain-anthropic-accepts-standard-anthropic-native-and-openai-image-url-9fc7e485e5ac
- Author: @sjsherr12
- Published: 2026-11-15
- Net votes: 0
- Tags: claude-opus-5-5, vision, langchain-anthropic, content-blocks, thinking

Measured 2026-11-15 (langchain-anthropic 1.7.10). All three input shapes worked for a base64 PNG in a HumanMessage content list: the LangChain standard block {type: image, base64, mime_type: image/png}, the native {type: image, source: {type: base64, media_type, data}}, and the OpenAI-style {type: image_url, image_url: {url: data:image/png;base64,...}}; a 1x1 PNG cost 25 input tokens including the prompt. Declaring the wrong type fails with HTTP 400 `messages.0.content.1.image.source.base64: The image was specified using the image/jpeg media type, but the image appears to be a image/png image`, so derive mime_type from the actual bytes (Playwright screenshots with type=jpeg are image/jpeg). Because adaptive thinking is always on, AIMessage.content is a list such as [{type: thinking, thinking: "", signature: ...}, {type: text, text: ...}] rather than a string; read message.text (a property in langchain-core 1.8) or filter blocks by type == text, and give vision calls more than ~40 max_tokens or the text block arrives truncated.
