258K can be the usable-window calculation 272,000 × 95% = 258,400, rather than the model's maximum capacity. Tibo announced a 500K default for GPT-6.1 Sol in Codex and Work; the same calculation would display 475K. Check the selected model, configuration overrides and current session before setting a window manually. Calculation · Tibo's announcement
Check these three things first
The selected model. The announcement concerns GPT-6.1 Sol. Check the model picker and, where supported, use /status to inspect the session. Record total capacity and usage separately: remaining context is not the total window. Status command
Existing overrides. Inspect ~/.codex/config.toml and trusted projects' .codex/config.toml, particularly model_context_window and model_auto_compact_token_limit. CLI flags, project configuration and a selected profile can take precedence over personal settings. In the CLI, /debug-config shows configuration layers and their sources. To follow the default, note and remove an old window override from the layer that sets it. Precedence · Diagnostics
Different sessions or model catalogs. Record the client version and compare a fresh chat with the existing one using the same model. LINUX DO users report different capacities in new and old chats; a GitHub issue also records one chat dropping from 475K to 258.4K. Catalog or cache inconsistency is worth investigating, but these reports establish neither a universal cause nor a minimum version guaranteed to fix it.
Set a 500K window manually
Put these entries at the top level of the configuration file that actually applies, before any [table] headers. Edit existing entries rather than adding duplicates. This example selects GPT-6.1 Sol; preserve your other settings.
model = "gpt-6.1-sol"
model_context_window = 500000model_auto_compact_token_limit controls when automatic compaction triggers; it does not enable a larger window. If you previously set it, note its value first. To return to the model's default threshold, remove that override and check higher-priority configuration layers. Official configuration reference
After saving, restart the client and inspect a fresh chat. This is a diagnostic step, not a guaranteed catalog fix. The window override is also capped by the model catalog's max_context_window; editing a local number cannot expand the server's capacity. Override implementation
FAQ
I set 500K but see 475K. Did it fail?
Not necessarily: 475K is exactly the result under the 95% calculation above. Existing context uses some of that capacity, so it is not a promise that you can paste another 475K tokens.
Is the missing 5% all reserved for compaction?
No: the source comment describes headroom for system prompts, tool overhead and model output. When neither model metadata nor configuration supplies a threshold, the cited version derives it at 90% of the window—450K for 500K; a lower catalog or configured limit can trigger compaction sooner, and this version-specific rule is not a permanent guarantee for every client. Threshold implementation
Why does the API list 1,050,000 while Codex defaults to 500K?
The API model specification, product default and catalog's maximum override describe different limits. Do not treat the API specification as the usable input budget of your current Codex session.
What if it still shows 258K after the change?
Record the version, model, configuration sources and capacities in new and existing chats for the relevant issue report; how the backend assigns defaults to a particular account remains uncertain. This article does not test 500K inference capacity, and the configuration example is not a test result.
Once the window is set
A larger window can retain more history; before writing, still check the current target, file state and permission. Read the 500K quick take and The plan was valid. The world had changed.