Claude cannot access images after performing a web search. TXT files seem to work fine, as does image access without prior search. Other models like o4 mini or o3 don't seem to have this issue.
Examples:
Claude 4 Sonnet (reasoning): https://kagi.com/assistant/157fe73b-d315-4ab4-8567-9c7399923c97
Claude 3.5 Haiku: https://kagi.com/assistant/6007e725-7f58-4c7c-a889-24daf8ece3df
In these examples, Sonnet appears to hallucinate tool calls and image content, while Haiku admits to not being able to see the image. This can easily be replicated using the same prompt with an image containing text that is preferably unrelated.
Test prompt used: "First, do research on diffusion models and give me a quick summary. Then, extract the text in the image and put it into a code block." (with a screenshot of lorem ipsum text)
Expected Behavior: Claude should be able to access and process images after performing web search, just as it does without prior search or when using other models. In this test case, Claude should extract the text from the image and output it in a code block.
