GPT-5.6 Sol effective context repeatedly flips between 828.4K and 258.4K in the same Codex App task

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

26.818.41509

Codex CLI bundled/available on the same machine: 0.149.0.

What subscription do you have?

ChatGPT subscription (tier omitted)

What platform is your computer?

Darwin 25.6.0 arm64 arm

What issue are you seeing?

In the same long-lived Codex App task, using the same model and local configuration, GPT-5.6 Sol's effective context window repeatedly changes between 828400 and 258400 tokens at turn boundaries.

This is not only a stale UI value. Across multiple days, the task's local rollout records contain token_count events whose model_context_window field repeatedly follows this pattern:

258400 -> 828400 -> 258400 -> 828400 -> 258400 -> 828400 -> 258400

A transition was also observed between adjacent turns without any model, client-version, or configuration change.

A fresh server-delivered model catalog currently returns:

gpt-5.6-sol:   context_window = 272000, max_context_window = 272000
gpt-5.6-terra: context_window = 272000, max_context_window = 872000
gpt-5.6-luna:  context_window = 272000, max_context_window = 872000
effective_context_window_percent = 95

The local configuration remains:

model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000

There is no model_catalog_json override, and I have not manually edited models_cache.json.

What steps can reproduce the bug?

The behavior is intermittent rather than deterministically triggered on every refresh:

1. Sign in to Codex App with ChatGPT authentication.

2. Configure GPT-5.6 Sol with:

   model = "gpt-5.6-sol"
   model_context_window = 1000000
   model_auto_compact_token_limit = 900000
   

3. Start or continue a GPT-5.6 Sol task.

4. Across turns and normal remote model-catalog refreshes, inspect the effective window shown by the App or the task's token_count.info.model_context_window events.

5. Observe the same task alternating between 828400 and 258400 without a local configuration or model change.

What is the expected behavior?

The server-delivered GPT-5.6 Sol context profile should remain stable for the same supported Codex App surface, model, client version, and configuration.

If the expanded context profile is controlled by a staged rollout, experiment, capacity mitigation, or product policy, a catalog refresh should not silently switch an active workflow between 872000 and 272000 maximum context profiles. The applicable eligibility and routing behavior should also be documented.

Additional information

The observations above were reproduced through the normal Codex App path without a custom model catalog.

Related reports:

• #39144 — Sol receives a 272K maximum while Terra and Luna receive 872K.
• #40106 — Sol remains capped at 272K despite the documented long-context configuration.
• #39102 — The bundled GPT-5.6 catalog was changed to allow context overrides up to 872K.

The distinguishing observation in this report is that the same task switched profiles repeatedly over time without any local model or configuration change. This suggests a time-varying rollout/routing component or inconsistent catalog propagation; that interpretation is an inference from the observed transitions.

The official GPT-5.6 Sol model page advertises a 1,050,000-token model context window. This issue concerns the Codex product's server-delivered catalog ceiling and its instability, not the underlying API model specification.

Start by inspecting the local rollout records and the token_count.info.model_context_window events, then trace how remote catalog refreshes feed the active task. Check the configuration and models_cache.json behavior described in the report. Done means the same supported task keeps a stable context profile, or the applicable rollout and routing behavior is documented.