OpenAINews

Illustration: nested grey window frames expanding toward a pale center on black.
Illustration

GPT-6.1 Sol default context jumps to 500k in Codex and Work

OpenAI Codex staff raised the default from 272k and stopped long-context fees on subscription usage.

If you run GPT-6.1 Sol in Codex or ChatGPT Work, you now get a 500k-token default context window instead of 272k. Use it for long agent runs that used to hit compaction mid-task. Subscription usage no longer pays long-context fees for Sol under this change.

What changed#

On Oct 11, 2026 at 7:45 PM ET, OpenAI's Tibo (@thsottiaux, bio: "Codex & ChatGPT @OpenAI") posted Day 7 of the Codex team's shipping streak: "We have increased default context length for GPT-6.1 Sol to 500k when used through Codex & Work, up from 272k or an ~ 80% increase." The full note says 500k was "carefully tested and is the new sweet spot for most of our users," that you can still override higher or lower, and that the maximum for GPT-6.1 Sol is 1M. A follow-up two minutes later: "With this change, we are not charging long context fees for GPT-6.1 Sol when drawing usage from the subscription."

Before: Sol defaulted to 272k in Codex and Work. Tibo's follow-up suggests long-context fees could previously apply to Sol subscription usage, though neither post states the old fee terms. After: 500k default, optional override up to 1M, no long-context fees for Sol when usage comes from the ChatGPT subscription.

As of this check (Oct 13 ~5:56 AM ET), the Codex changelog's newest dated product entry is still Oct 8 (Ultrafast). The change lives in staff posts, not the public changelog yet.

Why it matters#

Compaction summarizes earlier turns when a session nears the window. Tibo's stated goal is to "reduce occurrence of compactions" on tasks that need "a lot of context and nuance over periods of time." Fewer forced summaries means fewer lost decisions mid-refactor, mid-debug, or across a long Codex Goal run.

How to use it#

  1. Confirm you are on GPT-6.1 Sol in Codex or ChatGPT Work (not Chat). In the CLI: codex --model gpt-6.1-sol or /model.
  2. Audit overrides that can pin you to the old window. OpenAI's sample config documents model_context_window as "default: auto for model" and model_auto_compact_token_limit as "unset uses model defaults." Search your layers:
rg -n "model_context_window|model_auto_compact_token_limit" ~/.codex .codex
  1. If a value was only there to work around 272k, remove it and restart Codex so the new model default applies. Keep deliberate pins and document why.
  2. For a one-off larger window (up to 1M), set an override in ~/.codex/config.toml or pass --config for a single run. OpenAI's advanced config shows the key pattern:
model = "gpt-6.1-sol"
model_context_window = 1000000
  1. Start a fresh thread for the bigger default to apply cleanly.

Copyable prompt for a long multi-file job that benefits from the larger window:

You are on GPT-6.1 Sol with the new 500k default. Do not compact unless I ask. Walk the whole checkout for every call site of <SYMBOL>, list file:line, propose one minimal patch set, and keep the full call-site list in context until I say you are done.

Workflow people actually run#

Task: keep a multi-hour Codex coding thread coherent without mid-run summary loss. Tool: Codex or ChatGPT Work on GPT-6.1 Sol. Result: default window 500k (was 272k), fewer automatic compactions; subscription Sol usage without long-context fees (per Tibo). Cited: https://x.com/thsottiaux/status/2109430343346581527 and https://x.com/thsottiaux/status/2109430894604026039. Last checked: 2026-10-13.

Good to know#

  • Scope is Codex and Work only. Tibo did not claim Chat or the API product surface changed.
  • Built-in defaults sit at the bottom of Codex's config stack. CLI flags, project .codex/config.toml, profiles, user config, and managed defaults all outrank the new 500k if they set the same keys (see OpenAI advanced config / sample config).
  • Compaction still exists. A larger default delays it; /compact and a pinned model_auto_compact_token_limit still fire.
  • API list pricing for prompts beyond 272k is a separate surface. Do not assume the subscription fee waiver applies to API-key billing.
  • Watch next: an official Codex changelog or models-page line that states 500k; Tibo also said "We will go higher in the future." Separately, on Oct 12 at 9:03 PM ET he posted "Day 8/ Reset" and "Global reset landing by EOD PST time. Enjoy," then at 1:06 AM ET Oct 13: "What would you like with your side of reset? It's done."

Sources#

Last checked: 2026-10-13

Sources

  1. x.com
  2. x.com
  3. x.com
  4. x.com
  5. x.com
  6. developers.openai.com
  7. developers.openai.com
  8. learn.chatgpt.com

Something wrong? Tell the bot