[macOS app][GPT-6 Astra Ultra] Severe speed regression despite Fast mode being enabled

What version of the Codex App are you using (from the “About” dialog)?

26.903.61454 — released September 8, 2026.

What subscription do you have?

ChatGPT Pro — $200/month, 20x Codex usage tier.

What platform is your computer?

• macOS Tahoe 26.6.2
• MacBook Pro, 16-inch
• Apple M5 Max
• 128 GB memory

What issue are you seeing?

GPT-6 Astra Ultra has become dramatically slower than it was immediately after launch. This is not a small variation in task duration: the current execution speed is far below the launch baseline and has suddenly become slow enough to make normal Goal work difficult to use.

Fast mode was enabled for the affected run—the lightning/Fast indicator was active—but actual execution remained extremely slow.

In one captured Goal/Plan implementation run:

• Model: GPT-6 Astra Ultra
• Fast mode: enabled
• Permissions: Full access
• Subagents: 3 working
• At approximately 5m 31s, the task reported only about 10% progress and was still in early file-reading/command work.
• At approximately 5m 44s, it reported only about 20% progress and was still searching project files rather than completing the core implementation.

The displayed percentages are the task's own progress estimates, not a formal throughput benchmark, but the elapsed time, sparse completed work, and continuing early-stage file search document the practical slowdown. The app UI itself remained responsive; the problem is the model/agent execution pace, not a frozen renderer.

Earlier Astra Ultra runs on the same account and machine produced much faster visible progress. I do not have a launch-era server trace, so I cannot state the exact multiplier, but the regression is large and immediately noticeable.

The internal cause is unknown. Please check whether Fast was actually honored for this run, whether requests were queued or throttled, whether the run was silently routed to a slower/default service tier, or whether a recent model/backend/app change caused the regression.

What steps can reproduce the bug?

1. Open the Codex interface in the macOS ChatGPT desktop app.
2. Select GPT-6 Astra Ultra.
3. Enable Fast mode.
4. Start a normal multi-step coding Goal/Plan task in an existing project with Full access.
5. Observe that the main task and subagents advance far more slowly than Astra did near launch, despite Fast being active.
6. Submit in-app feedback with diagnostic logs enabled.

The exact slowdown may depend on backend routing or current service conditions, so this may not reproduce consistently on every account or region.

What is the expected behavior?

• Fast mode should provide materially faster and reasonably consistent execution than standard service.
• Astra Ultra should not suddenly become far slower than its own launch-period performance on the same account and machine without any explanation.
• If capacity, queueing, throttling, or service-tier fallback is affecting a run, the app should disclose that state instead of continuing to show Fast as active.
• The product should expose enough per-run information to distinguish model reasoning time, server queue time, tool execution, subagent scheduling, retries, safety review, and transport delays.

Requested investigation

Please correlate the feedback record below and inspect:

1. Whether the Fast service tier was actually applied to the main task and all three subagents.
2. Server-side queue/admission delay, time to first token, generation latency, and tool-call gaps.
3. Any silent service-tier fallback, throttling, overloaded routing, retries, or safety/guardian pauses.
4. Whether app version 26.903.61454 or a recent Astra backend/model update introduced a performance regression.
5. Why the UI provides no warning when effective performance is far below what the active Fast setting implies.

Please restore the earlier performance level or provide a clear explanation and visible degraded-service indicator when Fast capacity is unavailable.

Additional information

Feedback ID: 01a08711-c722-7b93-b166-3813ee781b1b

The in-app feedback submission completed successfully with diagnostic logs enabled.

The same run also produced a separate quota-accounting report in #44208. That issue concerns the abnormal weekly-quota deduction; this issue specifically concerns the severe execution-speed regression. The shared Feedback ID is intentional because both symptoms occurred during the same run.

No private source code, repository names, local paths, account identifiers, or device serial numbers are included in this report.

No files, tests, or code entry points are named. Start with feedback ID 01a08711-c722-7b93-b166-3813ee781b1b and its diagnostic logs, then compare Fast-tier routing, queue delay, first-token latency, generation latency, and tool-call gaps for the main task and three subagents. Done means determining whether routing, throttling, retries, or the app/backend versions explain the regression and whether degraded service is disclosed.