Why does Codex report a 258,400-token context window for GPT-5.6 Sol?

Codex
,
1

I’m using Codex Desktop / CLI 0.152.0 with ChatGPT authentication on macOS.

My config selects gpt-5.6-sol and does not override model_context_window or model_auto_compact_token_limit.

Observations:

  • Codex token-usage events consistently report model_context_window: 258400.
  • The bundled model catalog contains context_window: 272000 and max_context_window: 872000.
  • 258,400 is exactly 95% of 272,000.
  • Automatic compaction occurred at roughly 215K–243K input tokens in multiple sessions.
  • The API model page advertises a 1.05M context window and notes different pricing above 272K input tokens:
    GPT-5.6 Sol Model | OpenAI API
  • The Codex config reference documents model_context_window, model_auto_compact_token_limit, and model_auto_compact_token_limit_scope:
    Configuration Reference | ChatGPT Learn

Questions:

  1. Is 272K the intentional default context limit for ChatGPT-authenticated Codex?
  2. Is model_context_window = 872000 a supported way to enable the larger window, or would that merely make the client exceed a backend limit?
  3. Does using the larger window consume more Codex credits, corresponding to the API’s long-context pricing?
  4. Does model_auto_compact_token_limit_scope = "body_after_prefix" exclude the initial system/tool/instruction prefix, or only the retained prefix after the first compaction?
  5. Why is the Codex product limit substantially smaller than the model’s documented API context window?

I’d appreciate clarification from anyone familiar with the Codex client/backend behavior, especially whether the 272K default and 872K maximum are supported product limits or only internal catalog metadata.

2

Out of interest have you asked Codex this very question? It has access to the official docs.

3

Yes — these observations came from asking Codex to inspect its own live usage events, local configuration, and bundled model catalog, then comparing them with the official documentation.

The unresolved part is precisely what Codex itself cannot authoritatively confirm: whether the 272K default and 872K maximum are supported backend/product limits for ChatGPT-authenticated Codex, and how the larger window affects credits. I’m hoping someone from OpenAI or with direct knowledge of the Codex backend can clarify that distinction.

4

From my research:

Codex’s telemetry can show the effective context window and actual token usage, and backend model metadata may expose a default and override ceiling. However, those values alone do not confirm that 272K and 872K are supported product limits for all ChatGPT-authenticated accounts. The exposed usage API also does not provide the per-request credit calculation needed to establish the larger window’s charging impact. Those parts still need explicit OpenAI clarification.

So yes, OpenAI would need to chime in.

5

It doesn’t come enabled by default because it doesn’t necessarily improve the output, and it can greatly increase the rate at which usage is consumed.

Still, if you wish to enable it, here’s how to set it up: