[macOS / Pro] Spark limits show 100% remaining, but CLI 0.154.0 rejects gpt-5.3-codex-spark with HTTP 400

What version of Codex CLI is running?

Codex CLI 0.154.0. Also installed: Codex desktop 26.908.70816 (build 9275), with bundled CLI 0.154.0-alpha.6.2.

What subscription do you have?

ChatGPT Pro; authenticated in Codex with ChatGPT.

Which model were you using?

Explicitly requested gpt-5.3-codex-spark with diagnostic reasoning effort medium. This is the requested setting; no successful model execution or executed effort was established.

What platform is your computer?

macOS 27.0; Darwin 27.0.0 arm64 arm.

What terminal emulator and version are you using (if applicable)?

A noninteractive local CLI invocation from Codex desktop.

Codex doctor report

Not included. This report intentionally contains no diagnostic bundle, raw logs, source files or conversation history.

What issue are you seeing?

On September 15, 2026, the usage page showed 100% remaining for both the five-hour and weekly GPT-5.3-Codex-Spark usage limits. However:

• Spark was absent from the installed CLI model catalogs and the desktop task's exposed model-dispatch catalog.
• One isolated, ephemeral, read-only CLI request using the exact lowercase model ID failed before producing an assistant response.
• At 2026-09-16T03:32:00Z, the server returned HTTP 400, invalid_request_error:

The 'gpt-5.3-codex-spark' model is not supported when using Codex with a ChatGPT account.

The invocation ended with turn.failed and exit code 1. A missing-model-metadata warning appeared, but that warning does not establish that another model executed.

This report does not claim a fresh visual inspection of the desktop model picker. The catalog observations and CLI rejection were directly checked.

What steps can reproduce the bug?

1. Sign in to Codex with the affected ChatGPT Pro account.
2. Observe the separate Spark five-hour and weekly usage windows.
3. Inspect the available model catalogs.
4. From an isolated temporary working directory, run a text-only access probe with these options:

codex exec --ephemeral --ignore-user-config \
  --skip-git-repo-check --sandbox read-only \
  --model gpt-5.3-codex-spark \
  -c model_reasoning_effort=medium --json \
  "Do not use tools or inspect files. Reply exactly: SPARK_ACCESS_OK"

5. Observe the HTTP 400 rejection, turn.failed, exit code 1, and no assistant response.

Only one generation probe was performed. No client patch, downgrade, fabricated catalog, API-key switch or billing change was used.

What is the expected behavior?

If Spark is available to this account, a supported client should expose it and requests should succeed while quota remains. If access has ended or the model has retired, the quota UI and documentation should clearly state that status.

Please clarify current Spark availability, any effective retirement date, and whether a supported remedy exists for the quota/catalog/server mismatch.

Additional information

• Related report: #45594 describes the same mismatch. This separate report records my own macOS reproduction, feedback receipt and cropped quota evidence; it does not assert a different root cause.
• Verified feedback receipt: no-active-thread-01a0a84d-8ffd-70a3-b5e0-196097d39bcc. The official Codex feedback/upload API accepted this sanitized report with logs disabled and no conversation or attachments included.
• Support: Help Support confirmed escalation to a support specialist. No human diagnosis or case number was displayed.
• Documentation: The model guide, What's new, Codex changelog and launch announcement were reviewed. Current model documentation still lists Spark as a Pro preview. No formal Spark retirement notice was found in those pages or the 60 recent official GitHub release entries checked. This bounded search does not prove no announcement exists.
• Retirement reports: A community discussion reproduces an attributed September 11 announcement. The original social post could not be directly verified; retirement is not presented as the established cause.
• Separate subscription notice: The official Pro FAQ says new Pro $200 sign-ups/upgrades are paused from September 10, while existing Pro $200 subscriptions continue renewing normally. No causal connection between that pause and this Spark failure has been established.

Spark-only screenshot

The image contains only the Spark heading and its two usage windows. Billing information, other usage limits, notifications and surrounding content were cropped out.

Start with the documented codex exec invocation and compare the installed CLI model catalogs with the HTTP 400 response; review related report #45594 and the linked model documentation. Done means determining whether Spark is supported for ChatGPT Pro and identifying a concrete client, catalog, or documentation change, if the repository owns one.