Add Support for Claude-4.5-Sonnet & Gemini-3 in GAIA Workflow

[Lemaqwq]

Add Support for Claude-4.5-Sonnet & Gemini-3 in GAIA Workflow

Summary

When running run_gaia_workforce_claude.py with ModelType.CLAUDE_4_SONNET (claude-sonnet-4-5-20250929), two critical errors occur during task execution that prevent successful completion of GAIA benchmark tasks.

Environment

• Model: claude-sonnet-4-5-20250929 (Claude 4.5 Sonnet)
• OpenAI SDK: 2.15.0
• Framework: CAMEL-AI (modified 0.2.46)
• Python: 3.10

---

Error 1: Format Validation Error (JSON Parsing)

Symptom

The model returns plain text instead of valid JSON when structured output (response_format) is expected.

Log

2026-01-23 11:23:05,195 - camel.camel.models.openai_compatible_model - WARNING - Format validation error: 1 validation error for TaskResult
  Invalid JSON: expected ident at line 1 column 2 [type=json_invalid, input_value="I'll help you find Eliud...g for this information.", input_type=str]
    For further information visit https://errors.pydantic.dev/2.12/v/json_invalid. Attempting fallback with JSON format.

Analysis

The model response starts with "I'll help you find Eliud..." (conversational text) instead of a JSON object. This suggests:

• The response_format parameter may not be properly enforced for Claude models via OpenAI-compatible endpoint
• The beta.chat.completions.parse method behaves differently for Anthropic models

---

Error 2: Empty Content Block (BadRequestError 400)

Symptom

API returns 400 error stating text content blocks must be non-empty.

Log

openai.BadRequestError: Error code: 400 - {'error': {'message': 'messages: text content blocks must be non-empty (request id: 20260123195408753875355OP3qL5yB)', 'type': '<nil>', 'param': '', 'code': None}}

Stack Trace

File "/Users/lihengchen/Github/owl/camel/models/openai_compatible_model.py", line 235, in _arequest_parse
    return await self._async_client.beta.chat.completions.parse(
File ".../openai/resources/chat/completions/completions.py", line 1670, in parse
    return await self._post(
...
openai.BadRequestError: Error code: 400 - {'error': {'message': 'messages: text content blocks must be non-empty'...

Analysis

Somewhere in the message construction pipeline, an empty text content block is being sent to the API. This could be caused by:

1. Tool call results being processed incorrectly
2. Message history containing empty assistant responses
3. Incompatibility between OpenAI SDK v2.x beta.chat.completions.parse and Anthropic's API

---

Task Context

The error occurred during this GAIA benchmark task:

Worker node 5375996352 (A helpful assistant that can search the web, extract webpage content,
simulate browser actions, and retrieve relevant information.) get task e1fc63a2-da7a-432f-be78-7c4a95598703.0:
Search for Eliud Kipchoge's world record marathon time and determine his pace (speed in km/h or m/s).

Suggested Investigation

1. Check if beta.chat.completions.parse is the correct method for Anthropic models
2. Add validation to ensure no empty content blocks are sent
3. Verify response_format parameter compatibility with Claude 4.5 Sonnet
4. Consider using Anthropic's native SDK instead of OpenAI-compatible endpoint for structured outputs

---

Related Files

File	Line	Description
camel/models/openai_compatible_model.py	235	_arequest_parse method using beta.chat.completions.parse
camel/agents/chat_agent.py	839	_aget_model_response method
run_gaia_workforce_claude.py	-	Main entry point with model configuration

---

Reproduction Steps

1. Configure run_gaia_workforce_claude.py to use ModelType.CLAUDE_4_SONNET
2. Run the GAIA benchmark: python run_gaia_workforce_claude.py
3. Observe errors during web search tasks requiring structured output