Summary
In standalone ChatGPT Chat, GPT-5.6 Sol Thinking in both Extra High and Pro modes repeatedly attempts to retrieve external web resources through Python, requests, curl, or shell commands executed inside the offline computation sandbox.
Those calls inevitably fail because that environment has no outbound internet access or DNS resolution. ChatGPT should instead invoke its built-in web-search capability.
This report concerns standalone ChatGPT Chat only. It is unrelated to Codex, Work, the API, SDKs, custom GPT Actions, or locally executed commands.
Steps to reproduce
-
Start a new standalone ChatGPT conversation.
-
Select GPT-5.6 Sol Thinking with either Extra High or Pro.
-
Ask a research question that requires consulting external documentation, for example:
Apple’s newer generic helper-extension API. What is this? Can you dig deeper into it so I understand?
- Open the Activity panel and inspect the tool execution.
Actual behavior
ChatGPT decides to access Apple’s documentation by running Python through a shell command:
bash -lc python - <<'PY'
import requests
urls = [
'https://developer.apple.com/tutorials/data/documentation/xcode/creating-enhanced-security-helper-extensions.json',
'https://developer.apple.com/tutorials/data/documentation/extensionfoundation/building-an-app-extension-to-support-a-host-app.json',
'https://developer.apple.com/tutorials/data/documentation/servicemanagement/smappservice.json',
]
for u in urls:
r = requests.get(u, timeout=20)
print(u, r.status_code, r.headers.get('content-type'), len(r.content))
print(r.text[:200])
PY
The request fails because the execution environment cannot resolve external hostnames:
socket.gaierror: [Errno -3] Temporary failure in name resolution
This becomes:
urllib3.exceptions.NameResolutionError:
HTTPSConnection(host='developer.apple.com', port=443):
Failed to resolve 'developer.apple.com'
and finally:
requests.exceptions.ConnectionError:
HTTPSConnectionPool(host='developer.apple.com', port=443):
Max retries exceeded
The same general problem occurs when the model chooses curl, requests, urllib, or other networking clients. Switching libraries cannot solve it because all of them execute inside the same network-isolated environment.
In the attached example, the failed attempt consumed part of a 7 minute 28 second reasoning run. ChatGPT then continued, marked the task as completed, and displayed this activity statement:
Accessed Apple documentation JSON via curl
That statement is inaccurate. The displayed command failed before retrieving any data from developer.apple.com.
It then falls back to “Web search” (which works) but first attempting to do requests from an offline container is a waste of time and thinking resources. It happens very frequently.
Expected behavior
Ideally ChatGPT should be given access to the internet so that it can run these curl/python commands (for some file types it’s often more effective and actionable to do manipulate data with scripts than using web search).
If this is not possible, then for a task requiring external web retrieval ChatGPT should:
-
Invoke the built-in ChatGPT web-search or browsing tool.
-
Never attempt external DNS or HTTP access from the offline Python/shell sandbox.
-
If the web-search tool is unavailable, explicitly state that it is unavailable instead of trying equivalent networking libraries.
-
Detect the failed retrieval and avoid claiming that the documentation was successfully accessed.
-
Avoid producing conclusions that appear to rely on a source which the activity log shows was never retrieved.
Reproducibility
I have observed this repeatedly in both:
-
GPT-5.6 Sol Thinking — Extra High
-
GPT-5.6 Sol — Pro
Environment
-
Product: ChatGPT Pro 20x
-
Affected modes: GPT-5.6 Sol Thinking Extra High and Pro
-
Observed: 15 July 2026
-
Not involved: Codex, Work, API, SDK, or custom GPT Actions
Please route this to the team responsible for standalone ChatGPT tool selection and model-to-tool routing rather than the Codex or API teams.
