GPT-5.5 Fast spends long time in Thinking/Reading after latest Codex Desktop update

Codex app version

Desktop Codex: 26.623.42026 — released Jun 27, 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

019f0f86-0d61-77a1-b2c6-b4c5b7e270b9

Summary

Over the past few days, GPT-5.5 Fast in the Codex Desktop app has felt noticeably slower than before. The main symptom is not ordinary slow streaming speed: once output starts, streaming often looks relatively normal. The slowdown mostly happens before output starts, where Codex can remain in Thinking or Reading for a long time with no visible progress.

Observed behavior

• Start or continue a normal Codex task using GPT-5.5 Fast.
• Codex frequently remains in Thinking or Reading for a long time before producing visible output.
• Streaming speed after output begins appears mostly normal compared with the long pre-output delay.
• The machine itself remains responsive, and the Codex app UI is not completely frozen.
• The delay makes Fast mode feel much slower and less predictable for normal coding-agent workflows.

Possible contributing factor

After updating to the latest Codex Desktop build, I kept the app open continuously and did not quit, restart, or reboot it. I am not sure whether the slowdown is related to long-running app uptime after the update.

If this is related to keeping the Codex app open for a long time without restarting, then it still seems like an app optimization issue: long-running Codex Desktop sessions should not gradually become slower or more prone to long Thinking / Reading stalls over time.

Expected behavior

GPT-5.5 Fast should maintain fast pre-output responsiveness, or clearly show what it is waiting on when it spends a long time in Thinking / Reading before producing output.

Long-running app uptime should not cause Codex Desktop to become increasingly sluggish. If app state, cached sessions, background workers, resource accumulation, or reconnection state can degrade performance over time, Codex should either recover automatically or surface an actionable status to the user.

Impact

This makes GPT-5.5 Fast unreliable for day-to-day development workflows because the user cannot tell whether Codex is still actively working, reading files, waiting on something hidden, or silently stalled. It also makes the expected Fast-mode experience feel much slower than before.