Codex opens 1M-token GPT-5.6 Sol context for ChatGPT subscribers

• ChatGPT subscribers can now enable a one-million-token Codex session for GPT-5.6 Sol, a path that had previously worked only with API keys, according to thsottiaux's account update.
• The override selects GPT-5.6 Sol, sets a 1,000,000-token budget, and starts automatic history compaction at 900,000 tokens, as thsottiaux's configuration post documents.
• The larger budget delays compaction rather than removing it, leaving 100,000 tokens between the configured compaction threshold and the one-million-token ceiling in thsottiaux's setup.
• Tokens used beyond Codex's normal context window count 2x against usage limits, according to reach_vb's quota caveat.

The official Codex configuration reference puts durable user settings in ~/.codex/config.toml, while trusted repositories can supply project-scoped .codex/config.toml overrides. OpenAI's sample configuration lists the context-window and auto-compaction fields as optional manual model metadata, with model defaults used when they are unset.

The access change is narrow but useful: the one-million-token configuration now works for usage through ChatGPT accounts, not only API-key authentication.

The model's documented maximum is 1,050,000 tokens, while the published Codex configuration uses a round one-million-token budget. dkundel's confirmation separately described the option as available with a ChatGPT subscription.

The values belong at the top of ~/.codex/config.toml, before any TOML section headers. thsottiaux's configuration post also says to restart the Codex client and start a new session after saving.

• model selects GPT-5.6 Sol.
• model_context_window raises Codex's working context budget to one million tokens.
• model_auto_compact_token_limit tells Codex to compact history at 900,000 tokens.

A larger window retains more code, tool output, and conversation history before older material is summarized. The 900,000-token trigger means compaction begins at 90% of the configured budget, preserving 100,000 tokens of headroom.

The announcement still describes the normal Codex window as deliberately tuned for performance and cost. One user, aibuilderclub_, said they were retaining the default because compaction felt smooth and they did not want to spend usage on unused context.

The new switch changes quota accounting once a session crosses Codex's ordinary context window. reach_vb's quota caveat said those extra tokens count twice against usage limits; neither the configuration post nor that caveat states the default-window size or plan-specific caps.

That warning matches kevinkern's reaction that a million-token window fits edge cases more readily than daily work. The configuration is an opt-in override, while the ordinary window and automatic compaction remain the default behavior.

A persistent config edit is not required. The same values can be passed for one CLI invocation, leaving the saved default untouched:

OpenAI's Codex manual describes command-line overrides as appropriate for one-off situations. The persistent path still requires a client restart and a new session, per thsottiaux's original instructions.