[Bug]: Claude code will report an error when uploading pictures，API Error (500 {"error":{"message":"Error calling litellm.acompletion for non-Anthropic model: litellm.BadRequestError: OpenAIException - litellm.BadRequestError:

[Monkeycrop]

What happened?

Claude code will report an error when uploading pictures, but the same model, using the API base of other suppliers, is no problem

Relevant log output

API Error (500 {"error":{"message":"Error calling litellm.acompletion for non-Anthropic model: litellm.BadRequestError: OpenAIException - litellm.BadRequestError: VertexAIException BadRequestError - {\n  \"error\": {\n    \"code\": 400,\n    \"message\": \"Invalid value at 'contents[0].parts[1].inline_data.data' (TYPE_BYTES), Base64 decoding failed for \\\"{'type': 'base64', 'media_type': 'image/png', 'data': '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

Are you a ML Ops Team?

No

What LiteLLM version are you on ?

v1.75.5

Twitter / LinkedIn details

No response

[terrydang]

meet the same problem with you

[ineersa]

Same here

[ishaan-jaff]

pls give steps to repro - maybe a video will help

[ishaan-jaff]

does it happen with anthropic or non anthropic ?

[ineersa]

does it happen with anthropic or non anthropic ?

Non anthropic, local llama.cpp server.

Rolled back to older version

pip install "litellm[proxy]==1.75.4"

This one works as expected.

Steps to reproduce:

1. Use some llama.cpp model (GPT-OSS in my case).
2. Set it up in config:

litellm_settings:
  drop_params: true
model_list:
  - model_name: gpt-oss-120b
    litellm_params:
      # Tell LiteLLM to forward to an OpenAI-compatible backend
      model: openai/generic
      api_base: http://server:8052
      api_key: "EMPTY"
      max_tokens: 65536
      max_input_tokens: 65536
      max_output_tokens: 32768
      input_cost_per_token: 0
      output_cost_per_token: 0
      supports_function_calling: true
      supports_parallel_function_calling: false
      supports_response_schema: true
      supports_reasoning: true
      supports_tool_choice: true
      supports_web_search: true
  - model_name: "anthropic/*"
    litellm_params:
      # Tell LiteLLM to forward to an OpenAI-compatible backend
      model: openai/generic
      api_base: http://server:8052
      api_key: "EMPTY"
      max_tokens: 65536
      max_input_tokens: 65536
      max_output_tokens: 32768
      input_cost_per_token: 0
      output_cost_per_token: 0
      supports_function_calling: true
      supports_parallel_function_calling: false
      supports_response_schema: true
      supports_reasoning: true
      supports_tool_choice: true
      supports_web_search: true
router_settings:
  server_url: http://0.0.0.0:11434

3. Try to trigger builtin tool in claude:

What is 123456789 × 987654321? Please calculate it precisely.

Expected result (works on 1.75.4):

That's very strange, it works on 120b model on new version too, let me check small one.

[ineersa]

does it happen with anthropic or non anthropic ?

Checked small one, it works too, I guess you always have to have:

- model_name: "anthropic/*"
    litellm_params:
    .....

Inside LiteLLM config file, otherwise it will start to complain, works for me

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.