Codex on ChatGPT Pro: repeated 429s with only 4–5 total child agents across sessions

What subscription do you have?

ChatGPT Pro

Which model were you using?

GPT-5.6 Sol

What issue are you seeing?

I’m seeing a clear regression in Codex multi-agent concurrency on the same account after upgrading from Plus to Pro.

Before the upgrade, I could commonly run around 7–8 Codex sub-agents without this problem. Now the account repeatedly hits 429s with much lighter workloads.

What makes this especially suspicious is that the apparent limit seems to be account-wide, not per session:

• One Codex session with 3 child agents can work normally.
• If I open a second Codex session and add 2 more child agents, 429s can start appearing.
• In another controlled 4-child run, 1 finished and 3 eventually failed with 429 Too Many Requests after already making some progress.
• A 6-child run also reproduced the issue consistently.

So this does not look like a simple local per-session child-thread ceiling. It looks more like an account-level concurrency / burst / rate-control limit shared across Codex sessions.

I’ve also noticed a broader slowdown at the same time: new Codex sessions, existing sessions, and sub-agents all feel noticeably slower, with longer pauses between steps and slower reasoning/response progression.

Setup

I reproduced this on the direct official OpenAI/ChatGPT Codex path:

• one OpenAI account
• ChatGPT authentication
• no sub2api
• no multi-account routing
• no third-party model provider in the reproduction path

I understand that my local max_concurrent_threads_per_session = 8 setting is only a client-side ceiling and does not guarantee eight simultaneous upstream requests.

The issue is the change in behavior on the same Codex account: workloads that were previously routine are now being throttled after upgrading to Pro.

Expected behavior

A ChatGPT Pro Codex account should not regress from routinely handling ~7–8 sub-agents to repeatedly hitting 429s with only around 4–5 total child agents across sessions, unless there is a documented server-side limit or some other account-level policy that explains the change.

Additional information

OpenAI Support Case #14001008 is already open. Support has confirmed that my overall Codex usage/spend limit is not exhausted, but they have not been able to identify the specific backend limiter or verify whether the account is in a reduced concurrency/burst bucket.

Could the Codex team please check whether there is:

• a recent regression in Pro concurrency/burst handling,
• an unexpectedly low account-level rate-control bucket,
• an account-specific throttling state,
• or a configuration/entitlement issue introduced around the Plus → Pro upgrade?

I have exact timestamps and runtime request IDs for multiple reproductions and can provide them privately if a maintainer wants to correlate them with server-side traces.

No repository file, test, or entry point is identified; the report concerns account-level Codex throttling across sessions. Start with the supplied reproduction timestamps and runtime request IDs, alongside Support Case #14001008, and correlate them with server-side traces. Done would mean identifying the limiter or entitlement regression and documenting the confirmed behavior.