[Windows][Regression] Codex spends 90+ minutes in Thinking; 7 minutes for Hello (Sol Max and Astra, Plus) #53281

Summary

Over approximately the past two weeks (late September–October 12, 2026), Codex on Windows has shown a severe latency/stalling regression. Both GPT-6.1 Sol Max and Astra are affected. Even trivial conversational replies can take minutes, while code-analysis turns may stay in “Thinking” for more than 90 minutes without a final result.

This report concerns response latency and agent progress, not a request for the model to solve complex tasks instantly.

Environment

• Subscription: ChatGPT Plus
• Client: Codex desktop app on Windows
• Models: GPT-6.1 Sol Max; Astra (similar symptoms observed)
• App/build version: Not captured yet
• Period observed: Approximately two weeks, including October 12, 2026
• Frequency: Repeated, across different tasks

Actual behavior / observed incidents

Incident A — 1h32m44s “Thinking” during read-only source analysis

The prompt asked Codex to analyze an existing quest dependency chain, identify a suitable alternative branch, check level requirements, and identify where it could return to the main storyline. No file modifications were requested.

• 50m46s: Still shown as working / “Thinking”; no useful answer.
• 55m34s: Two completed tool commands visible: a rg -n search and a Python invocation reading/analyzing JSON project data.
• 1h32m44s: The same two tool commands were visible; no additional tool activity was displayed and no final actionable response had appeared.
• The visible timeline therefore showed ~37 minutes with no further visible progress, and 92+ minutes without completion.

This does not prove no internal computation was happening; the issue is that the agent provided neither a result nor meaningful progress/timeout feedback.

Incident B — ~7 minutes for “Привет” (“Hello”)

A simple greeting took approximately seven minutes to receive a response. Such latency is disproportionate to the request.

Incident C — slow acknowledgment after successful file test

After Codex reported preparing a corrected game-cinematic file, I replied only “Работает^^” (“It works”). Codex entered “Thinking” again, with the timer at 55 seconds and no visible answer when captured.

Incident D — Astra also affected

Similar significant slowdowns and lengthy “Thinking” states have occurred on Astra over approximately the same two-week period. This does not appear isolated to one Sol Max configuration.

Expected behavior

• Trivial conversational turns should be acknowledged promptly.
• Tool-backed source analysis should either make timely progress or explain the blocker.
• A stalled turn should be surfaced as such, with a timeout/retry/recovery path.
• The UI should distinguish inference, queue waiting, active tool calls, and agent stalls where possible.
• The agent should not remain active for 90+ minutes without an actionable output or useful progress information.

Reproduction

1. Launch Codex desktop on Windows using a Plus account.
2. Select GPT-6.1 Sol Max and ask it to analyze a sizeable existing source-data dependency graph (read-only task).
3. Observe the tool activity and “Thinking” elapsed-time display.
4. Send simple conversational messages such as “Hello” or a short test confirmation and measure latency.
5. Repeat on Astra.

The issue appears intermittent and may not reproduce on every attempt.

Impact

This makes the tool unreliable for real development workflows, especially when interactive iteration and debugging are required. Extended stalls waste working time and make usage under subscription limits unpredictable.

Evidence

Screenshots were captured at 50m46s, 55m34s, and 1h32m44s during the same stalled task, plus a screenshot showing 55s of thinking after “Работает^^”. Images can be provided upon request (the issue is being filed via an integration that does not support attaching images).

Request

Please investigate potential regressions in model routing/inference scheduling, context processing, and agent execution between tool calls, especially on Windows desktop and for GPT-6.1 Sol Max/Astra. Please also improve stall detection, timeouts, and progress visibility.

A potentially related Windows stall report is #52694, but I cannot confirm it shares the same root cause.

I am reporting observed symptoms, not asserting intentional throttling by subscription tier.