GPT-5.5 Fast session freezes shortly after start, then resumes with extremely slow streaming

Codex app version

Desktop Codex: 26.609.41114 — released Jun 12, 2026

Platform / plan

• macOS Tahoe 26.5.1
• MacBook Pro, 16-inch
• Apple M5 Max
• 128 GB memory
• Pro plan
• Mode: GPT-5.5 Fast

Related feedback ID

019ec46c-d057-7253-9bf1-d64c1e8a9f07

Summary

GPT-5.5 Fast is again showing a serious session-freeze and slow-streaming issue.

Shortly after starting a new task, the Codex session suddenly became stuck. The Codex app itself did not freeze, and my computer was not frozen. Only the active Codex conversation appeared to be stuck.

During the stuck period, the session did not show Thinking, and it was not running any visible terminal command. It simply stopped progressing.

After that, the task resumed with extremely slow streaming output.

Observed behavior

• Start a Codex task using GPT-5.5 Fast.
• Shortly after the task begins, the conversation suddenly stops progressing.
• The app itself remains responsive.
• The computer remains responsive.
• The conversation does not show Thinking.
• No terminal command is visibly running.
• The task appears stuck with no clear status.
• Later, streaming resumes, but the output becomes extremely slow.

Expected behavior

GPT-5.5 Fast should not silently freeze a conversation shortly after a task starts.

If Codex is thinking, reading files, waiting on tools, running a command, or reconnecting, the UI should clearly show that state. If the task is blocked, Codex should surface an actionable status or error instead of leaving the conversation idle with no visible progress.

Fast mode should also stream at a speed consistent with the expected Fast experience.

Actual behavior

The conversation can silently freeze shortly after starting. The UI does not show Thinking, no terminal is running, and there is no clear explanation. After the freeze, the response stream becomes extremely slow.

Why this matters

This is a serious reliability and performance problem. The user cannot tell whether Codex is still working, stuck, waiting on a hidden operation, or failing silently.

It also wastes time and usage because the task may sit idle without progress, then continue with a degraded streaming speed. This makes GPT-5.5 Fast unreliable for normal development workflows.

Regression note

This issue seemed to temporarily stop happening after a recent update about a day ago, but it has now reappeared in the latest build.

Request

Please investigate why GPT-5.5 Fast sessions can silently freeze shortly after task start and then resume with very slow streaming.

Codex should:

• show a clear task state when it is thinking, reading, running tools, or waiting,
• avoid silent conversation freezes,
• recover cleanly if the stream stalls,
• and maintain expected Fast-mode streaming performance.