Skip to content

gpt-5.6-sol repeatedly identifies itself as GPT-5.4 across official Codex harness & Pi sessions #33502

Description

@hhushhas

What version of Codex CLI is running?

codex-cli 0.144.2 (clean reproduction); 0.144.1 (historical session)

What subscription do you have?

ChatGPT Pro (20x); verified for Trusted Access for Cyber (TAC)

Which model were you using?

gpt-5.6-sol

What platform is your computer?

Darwin 25.5.0 arm64 arm

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

Clean repro: Zed 1.10.3 stable integrated terminal, zsh. Historical: Codex TUI via VS Code, zsh.

Codex doctor report

Not available: `codex doctor --json` produced no output and did not complete within 30 seconds.

What issue are you seeing?

With gpt-5.6-sol explicitly selected, official Codex sessions repeatedly identify the model as GPT-5.4.

Incident 1: fresh controlled reproduction

Codex CLI 0.144.2 was invoked with --model gpt-5.6-sol in an empty temporary directory. No tools were used. The assistant first answered “I’m GPT-5.” When asked for the exact version, it answered “GPT-5.4”.

  • Thread: 019f691c-3c6f-7051-ac65-27cf0e2b2556
  • Relevant turn: 019f691c-c930-7731-a801-3d6641fc38b1
  • Assistant message item: msg_02410fa0e7b273df016a58594e39a48191919e880f3e36ee62
  • UTC: 2026-07-16T04:08:46.357Z

Incident 2: historical Codex TUI session

The saved rollout contains 12 turn_context records, all selecting gpt-5.6-sol. During one turn, the agent generated four actual tool calls using complain -m gpt-5.4.

  • Thread: 019f604c-c452-7ef0-b6fd-8cd9e9987b09
  • Relevant turn: 019f607e-01e2-72a0-813f-8babe95be944
  • Example reasoning item: rs_0440faa65d4d9359016a562b9bed7c8195b7461b8bcca003c0
  • Example tool item: ctc_0440faa65d4d9359016a562b9c54648195b888685b62da1ad4
  • UTC: 2026-07-14T12:29:22.460Z

I understand model self-identification is not authoritative proof of the backend route. I am reporting the mismatch and providing server-correlatable identifiers so OpenAI can verify which model actually served these requests.

What steps can reproduce the bug?

Uploaded threads: 019f691c-3c6f-7051-ac65-27cf0e2b2556 and 019f604c-c452-7ef0-b6fd-8cd9e9987b09

What is the expected behavior?

When gpt-5.6-sol is selected, the served model should match that selection. If the model cannot reliably know its exact identity, it should avoid asserting an incompatible version. OpenAI should confirm the actual backend route for both uploaded threads.

Additional information

I am on ChatGPT Pro (20x) and verified for Trusted Access for Cyber (TAC).

Sanitized clean reproduction archive:
https://github.com/hhushhas/openai-model-routing-reproductions/releases/tag/codex-clean-repro-2026-07-16

A broader local audit found 77 unique model-label mismatch occurrences across 40 official Codex threads: each saved turn_context requested gpt-5.6-sol, while the actual generated tool call used complain -m gpt-5.4. These are verified client-side mismatches, not proof of backend routing; the request is for OpenAI to verify the serving model from server logs.

Uploaded threads: 019f691c-3c6f-7051-ac65-27cf0e2b2556 and 019f604c-c452-7ef0-b6fd-8cd9e9987b09

Metadata

Metadata

Assignees

No one assigned

    Labels

    CLIIssues related to the Codex CLIbugSomething isn't workingmodel-behaviorIssues related to behaviors exhibited by the model

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions