[Model Identity Mismatch] Session metadata shows gpt-6-astra but assistant self-reports as GPT-5.5

What issue are you seeing?

The model selector is set to gpt-6-astra, but the assistant identifies itself as
OpenAI GPT-5.5. All authoritative session metadata fields confirm the effective
model is gpt-6-astra, so this is not a user selection error. Task performance is
also noticeably degraded, and reasoning_output_tokens is 0 for a simple identity
question at medium effort.

Evidence from the session:

• session_id: 01a09b45-a73b-77a2-bb4a-479cda10d3e0
• turn_id: 01a09b47-0f51-7e13-ad1f-3a7f0913232a
• CLI version: 0.154.0-alpha.6.2
• Source: vscode (Codex desktop)
• Plan: pro
• Model provider: openai

Authoritative model fields (all say gpt-6-astra):

• turn_context.payload.model: gpt-6-astra
• world_state.state.model: gpt-6-astra
• world_state.state.collaboration_mode.model: gpt-6-astra
• world_state.state.personality.model: gpt-6-astra

Assistant self-report:
"I'm ChatGPT, running as an OpenAI GPT-5.5 model in this conversation."

Performance / token usage:

• input_tokens: 27997
• output_tokens: 24
• reasoning_output_tokens: 0
• time_to_first_token_ms: 2966
• duration_ms: 4514

What steps can reproduce the bug?

1. Open Codex in the ChatGPT desktop app.

2. Select gpt-6-astra from the model picker (reasoning_effort: medium).

3. Send the following prompt:

   "What AI model are you? Please answer me seriously, and do not deceive me in any way."

4. Observe the assistant replies:

   "I'm ChatGPT, running as an OpenAI GPT-5.5 model in this conversation."

What is the expected behavior?

The assistant should identify itself consistently with the authoritative session
metadata (gpt-6-astra), or state that it cannot reliably introspect its own serving
model. The effective model should match the user's explicit selection, and reasoning
token usage should be consistent with gpt-6-astra at medium effort.

Additional information

This matches a previously reported pattern in #36872, where users selecting a newer
model also saw the assistant self-report as GPT-5.5 despite session metadata
confirming the correct model.

The sanitized rollout log is attached. No PII is included.

rollout-2026-09-13T07-57-08-01a09b45-a73b-77a2-bb4a-479cda10d3e0(After desensitization).txt

Start by reviewing the attached sanitized rollout log and the related report in #36872, comparing the authoritative model fields with the assistant's self-report and token usage. Done means identifying the source of the mismatch and ensuring the response is consistent with the effective model or clearly states that model introspection is unreliable.