Users report GPT 5.6 Sol restores its 1M-token context window
Users report GPT 5.6 Sol's selected 1M-token context window initially reset to roughly 258K after a message. A later user confirmation said the 1M setting was restored.
TL;DR
- The 1M-token configuration could appear in Codex, then collapse to 258K after the first message, as LLMJunky's report documented.
- A later test says the setting now persists, although LLMJunky's confirmation displays an 828K working capacity rather than a literal 1M.
- An independent ChatGPT-subscription test reproduced the 828K ceiling that LLMJunky's confirmation shows after a 1M request.
- The point at which a long session starts consuming extra usage remains unspecified, according to LLMJunky's follow-up.
A ChatGPT subscription test found that a one-million-token request topped out at 828K. A mirror of Tibo Sottiaux's instructions reports that the configuration now works on ChatGPT accounts, where it had previously been limited to API-key use.
The setting that reverted
The local configuration posted by LLMJunky asked Codex for a one-million-token window and an 800K auto-compaction limit:
model = "gpt-5.6-sol"model_context_window = 1000000model_auto_compact_token_limit = 800000
The reported failure happened after a message: /status moved from zero used out of 1M to 22.5K used out of 258K. LLMJunky said it was a new session, then corrected an earlier objection to say 1M was not active for subscription users of GPT-5.5 or GPT-5.6 at that time in a follow-up.
The 828K working ceiling
The working screenshot reports 36K tokens used out of 828K. The subscription test independently logged the same ceiling after specifying 1M, with the ordinary subscription default at 258K.
Tibo Sottiaux, an OpenAI Codex engineer, said in a mirror of his original post that the setting now works for ChatGPT accounts and that the default was tuned for performance and cost. Neither account explains why the exposed capacity stops at 828K.
Usage threshold
The higher 372K cap had launched, then been reduced because of a bug, LLMJunky wrote. When asked whether extra usage begins at 372K or 272K, LLMJunky's question surfaced the split and their response said the cutoff was unclear.
The same user later warned that the larger window consumes more usage and could still be buggy in a follow-up.