GPT-5.6 Sol and Terra exit OpenClaw scheduled research tasks before calling web search; GPT-5.5 completes the same workflow

What issue are you seeing?

Summary

After switching an existing unattended OpenClaw research workflow from GPT-5.5 to GPT-5.6 Terra and GPT-5.6 Sol, both GPT-5.6 models exit before invoking the available web search tools.

The same workflow completes successfully and consistently with GPT-5.5.

Environment

• Device: Mac mini M4
• Operating system: macOS
• Integration: OpenClaw using Codex models
• OpenClaw version: Latest beta supporting GPT-5.6 on July 10, 2026
• Models tested:
• GPT-5.5
• GPT-5.6 Terra
• GPT-5.6 Sol
• Workflow type: Unattended scheduled research
• Network access and search tools: Enabled

Workflow

The workflow generates three daily reports:

1. AI news
2. International financial news
3. Middle East situation report

The model is expected to:

1. Search for current information
2. Verify sources
3. Select four to five important events
4. Generate a report using a predefined format

Source discovery is part of the task, so no fixed RSS feed or prebuilt source package is provided.

Actual behavior

GPT-5.6 Terra was tested three times. All three runs exited before calling web search and returned a response equivalent to:

Unable to verify. Insufficient information.

GPT-5.6 Sol was then tested three times and produced the same result.

Neither GPT-5.6 model attempted to invoke web search.

GPT-5.5 was tested three times with the same workflow and completed all three reports successfully.

Controlled A/B test

All tests used:

• The same device
• The same OpenClaw environment
• The same runner
• The same prompt
• The same tools
• The same tool permissions
• The same network configuration
• The same workflow

The only changed variable was the model.

| Model | Runs | Search invoked | Reports completed |
|---|---:|---:|---:|
| GPT-5.5 | 3 | 3 | 3 |
| GPT-5.6 Terra | 3 | 0 | 0 |
| GPT-5.6 Sol | 3 | 0 | 0 |

Expected behavior

When search tools are available and the workflow requires current research, the model should attempt to search and verify information before deciding that the task cannot be completed.

The absence of a prebuilt source package should not cause the model to exit before attempting the available tools.

Impact

This causes silent failures in unattended production workflows.

The scheduled process technically finishes normally, but the business task is not completed. Monitoring based only on process completion may incorrectly mark the run as successful.

Workaround

The three scheduled research workflows have been pinned back to GPT-5.5.

GPT-5.6 continues to work well for coding, debugging, and interactive analysis in the same environment.

I can provide redacted prompts and execution logs if needed.

What steps can reproduce the bug?

1. Configure an OpenClaw scheduled research workflow with web search enabled.

1. Use a prompt that instructs the agent to:

• Search for current information
• Verify sources
• Select four to five important events
• Generate a report in a predefined format

1. Run the workflow three times with GPT-5.5.

1. Confirm that GPT-5.5 invokes web search and completes the report each time.

1. Keep the same device, OpenClaw environment, runner, prompt, tools, tool permissions, network configuration, and workflow.

1. Change only the model to GPT-5.6 Terra.

1. Run the workflow three times.

1. Confirm that GPT-5.6 Terra exits before invoking web search and returns “Unable to verify” or “Insufficient information.”

1. Change only the model to GPT-5.6 Sol and repeat the test three times.

1. Confirm that GPT-5.6 Sol produces the same early-exit behavior.

What is the expected behavior?

When web search tools are available and the scheduled workflow explicitly requires current research, the model should attempt to invoke web search, verify sources, and complete the report.

If the search tool fails, the model should report the tool failure or retry.

The absence of a fixed RSS feed or prebuilt source package should not cause the model to exit before attempting the available search tools.

The behavior should remain compatible with the same workflow that completes successfully with GPT-5.5.

Additional information

_No response_