Severe streaming slowdown in Codex: Fast mode outputs only a few characters every several seconds and then stalls

Codex app version

Desktop Codex: 26.608.12217 — released Jun 8, 2026

Platform / plan

• macOS Tahoe 26.5.1
• Apple Silicon MacBook Pro
• Pro plan
• Model/speed: GPT-5.5 with Fast mode enabled

Related feedback ID

019eb3e5-1aa2-7433-95d3-2c39514a846b

Summary

Codex streaming became extremely slow today, even with GPT-5.5 Fast enabled. This is a serious performance issue because the response stream is progressing at an abnormally slow pace and heavily impacts task speed and efficiency.

I recorded a screen video as evidence. The recording clearly shows that the streamed output only advances by a few characters every several seconds. This pace is not normal for Fast mode and makes the task feel nearly unusable.

Observed behavior

During an active Codex task:

• Streaming output is extremely slow.
• Only a few characters appear every several seconds.
• The delay is visible and persistent, not just a brief momentary pause.
• This happens despite using GPT-5.5 Fast / fast-speed mode.
• Later in the same task, Codex reached a Reading ... step and then appeared to stall there.

Expected behavior

Fast mode should provide noticeably faster streaming and task progress. Even if the task itself requires reasoning or code inspection, the visible output stream should not update only a few characters every several seconds.

A Reading step should also make progress or provide a clear indication if it is blocked, instead of appearing stalled.

Actual behavior

The stream is much slower than expected for Fast mode. Output appears only in tiny fragments after long pauses, and the task later stalls at a reading step.

Why this matters

This severely reduces Codex's usefulness for normal development workflows. Long-running tasks already take time, but when the stream itself becomes this slow, users cannot efficiently monitor progress, understand what Codex is doing, or continue work at a reasonable pace.

This also increases the perceived cost of tasks because the user spends much more time waiting for output, even when using a mode that is expected to be faster.

Request

Please investigate whether there is a streaming, backend latency, throttling, or Fast-mode performance regression in the latest Codex build.

It would also help if Codex could expose clearer status when the model is slow due to backend queueing, streaming delay, throttling, or a stuck Reading operation, so users can distinguish normal work from a degraded or stalled session.

https://github.com/user-attachments/assets/bac7f740-565b-409b-a297-dfaec54ca87f