GPT-5.5 reports 258400 context window in Codex despite published 400K window

Summary

Codex reports GPT-5.5 as having a 258400 token context window in live session events, while OpenAI's GPT-5.5 launch page says GPT-5.5 in Codex has a 400K context window.

Official source: https://openai.com/index/introducing-gpt-5-5/

Environment

• Codex CLI: 0.124.0
• Model: gpt-5.5
• Plan: Pro
• Platform: macOS

Observed

Live session JSONL events contain:

{"type":"task_started","model_context_window":258400}

and token count events also report:

{"type":"token_count","info":{"model_context_window":258400}}

The context-window-size status line showed about 258k for GPT-5.5 after a turn, even though it had previously shown 400k and there is a separate available-context metric.

Local evidence

codex debug models reported this before a local hotfix:

{
  "slug": "gpt-5.5",
  "context_window": 272000,
  "max_context_window": 272000,
  "effective_context_window_percent": 95
}

That exactly explains the emitted window:

272000 * 0.95 = 258400

Local config already specified:

model = "gpt-5.5"
model_context_window = 400000
model_auto_compact_token_limit = 370000

Expected

For GPT-5.5 in Codex, the total model context window displayed and emitted as model_context_window should be 400000, matching the published Codex window. Available/remaining context should be shown separately and may change with usage.

Suspected cause

The Codex model catalog appears to encode 272000 as the total GPT-5.5 Codex context window and then applies the 95% effective-window multiplier, producing 258400. The 272000 value looks like the input portion of a 400K split (272k input + 128k output) rather than the full Codex model window.