On August 17, OpenAI Codex engineer Tibo Sottiaux posted on X the exact steps to manually override Codex’s default context cap and turn on GPT-5.6 Sol’s full 1,000,000-token window — a setting that, until that post, was effectively only available to developers paying through API keys. ChatGPT Plus and Pro subscribers using the Codex CLI had no documented way to reach it. Now they do. It takes editing a config file by hand, and it comes with a catch two open GitHub issues have been documenting for over a month: the number Codex actually gives you by default is a fraction of what OpenAI advertises.

Quick Summary: What Changed

Detail	Info
Date	August 17, 2026
Who posted it	Tibo Sottiaux, OpenAI Codex engineer, via X
What changed	Manual config.toml steps to enable GPT-5.6 Sol’s full 1M-token context in Codex, now confirmed to work on ChatGPT Plus/Pro accounts, not just API keys
Advertised context window	1,050,000 tokens (GPT-5.6 Sol’s published model spec)
Documented default in Codex	372,000 tokens raw / 353,400 effective, per issue #31860 — later cut to 272,000 raw / 258,400 effective, per issue #32806
Gap vs. spec	The post-regression default is roughly 25% of the advertised window
The fix	Manually edit ~/.codex/config.toml — most users never find this

Bottom line: OpenAI’s own engineer had to hand you a workaround for a limit OpenAI’s own product imposes. The 1.05M number on GPT-5.6 Sol’s spec sheet was never the number most Codex users were actually running on.

---

What Actually Happened

Sottiaux’s post is short and unambiguous. Add three lines to the top of ~/.codex/config.toml, before any [section] header:

model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000

Save it, restart Codex, start a new session. The first line pins the model. The second tells Codex to budget for a full million-token window instead of whatever it was defaulting to. The third sets automatic history compaction to kick in at 900,000 tokens, leaving roughly 100K tokens of headroom before Codex has to start summarizing older context to make room for new.

Sottiaux framed it as a documentation gap, not a new feature: “Even though we have tuned the context limit in Codex to be set optimally when it comes to performance and cost, this is a common ask, so here it is documented.” Translation — the override already worked technically. OpenAI just hadn’t told anyone how to use it, and until this post, reporting suggests it only reliably worked for API-key billing, not subscription accounts on Plus or Pro. That’s the actual news here: a capability that existed quietly for API users got confirmed and documented for the much larger population of people paying for Codex through a ChatGPT subscription.

How Do You Enable Codex’s Full 1M Context Window?

1. Open ~/.codex/config.toml in a text editor. Create the file if it doesn’t already exist.
2. Add three top-level keys before any [section] header: model = "gpt-5.6-sol", model_context_window = 1000000, and model_auto_compact_token_limit = 900000.
3. Save the file and fully restart Codex. The override won’t apply to a session that’s already running.
4. Start a new session rather than continuing an existing one — Codex reads the context budget at session start.
5. Verify it took effect. Reported behavior on subscription billing has been inconsistent enough that checking your actual effective context during a session is worth the extra step, rather than assuming the config value is being honored.

That last step matters more than it sounds. It’s the same inconsistency that makes this whole story worth writing about.

The Gap Between the Spec Sheet and the Product

GPT-5.6 Sol’s published model specification lists a 1,050,000-token context window — the number OpenAI puts in its own documentation, and the number we’ve cited in earlier coverage comparing GPT-5.6 Sol against Grok 4.6 and Claude Fable 5. That’s the spec. It is not what Codex, the product built to run that model, was actually giving most users by default.

GitHub issue #31860, opened July 9, documents the first version of the gap. The reporter found Codex’s model catalog capping GPT-5.6 Sol at a raw 372,000-token context window, with a 95% effective-window policy bringing usable context down to 353,400 tokens. Do the same math against the advertised 1.05M spec and you’d expect roughly 997,500 usable tokens. Instead, Codex users were getting about 35% of that.

It got worse from there. Issue #32806 documents a second, undocumented cut on July 13 — the same window dropping again, from 353,400 effective tokens down to 258,400. Raw capacity went from 372,000 to 272,000 tokens. No changelog entry, no announcement. Users noticed because their sessions started compacting and losing history earlier than they had the day before, and started comparing notes in the issue thread. The reporter measured the transition happening live, noting that “old and new values briefly overlapped across existing and newly started threads, consistent with a phased or cached server-side catalog rollout” — the kind of detail that only shows up when someone is actively monitoring their own token budget because the product stopped behaving the way it did the week before.

Put the two numbers next to the spec sheet and the picture is stark:

Stage	Raw context window	Effective context window	% of advertised 1.05M spec
Advertised spec	1,050,000	~997,500 (at 95% policy)	100%
Codex default, per #31860	372,000	353,400	~35%
Codex default, per #32806 (later)	272,000	258,400	~26%
After manual config.toml override	1,000,000	up to 900,000 (compaction threshold)	~90-95%

That’s not a rounding difference or a conservative default tuned for cost. That’s most Codex users running long agentic sessions on roughly a quarter of the context window OpenAI advertises for the model doing the work, without any indication in the product that a much larger window exists and is one config edit away.

Why This Matters

Context window size isn’t a spec-sheet vanity metric for a tool like Codex. It’s the ceiling on how much of your actual codebase, terminal output, and conversation history the model can hold in its head before it starts forgetting or summarizing. A smaller effective window means more frequent compaction, and compaction means the model is working from a compressed summary of what happened earlier in the session instead of the real thing. For a quick single-file fix, that rarely matters. For a long agentic run — refactoring across a large repo, debugging a chain of failures, anything that stretches past a few dozen tool calls — it’s the difference between Codex remembering the constraint you set 40 steps ago and re-deriving it wrong.

We’ve flagged context window as a real differentiator before, not a marketing footnote. In our comparison of Grok 4.6, GPT-5.6 Sol, and Claude Fable 5, GPT-5.6 Sol’s 1.05M-token spec was one of its clearest advantages over Grok 4.6’s 500K window. That comparison used the number OpenAI publishes. If Codex users were actually capped at 258,400 tokens for over a month without anyone outside a GitHub issue thread knowing it, the advantage on paper and the advantage in practice were two different things — and the gap only closes if you know to go find config.toml and edit it yourself.

There’s also a cost dimension the config steps don’t advertise loudly. Sottiaux’s own post concedes the point directly: Codex’s default is “tuned optimally when it comes to performance and cost.” A bigger context budget means more tokens processed per turn, and for anyone on usage-based billing or a metered weekly quota, manually opting into the full 1M window will burn through allowance meaningfully faster than the tuned default. The workaround isn’t free. It’s a tradeoff OpenAI’s own engineer flagged in the same breath as the fix, and it’s easy to skip past when you’re just excited the setting finally exists.

What Are Your Options Now

If you’re running long agentic sessions in Codex, check your actual context budget first. Don’t assume you’re getting anywhere close to 1.05M tokens by default — the GitHub issue history says you’re probably not, and the exact number has moved twice already without notice.

Apply the config.toml override if long-context work is a regular part of your workflow. The three-line edit is straightforward, but restart Codex fully and start a fresh session — the override doesn’t apply retroactively to a session already in progress.

Budget for the cost tradeoff before you flip it on for everything. A 1M-token context budget processes more tokens per turn than Codex’s tuned default. If you’re on a metered plan, reserve the override for sessions that actually need the extra headroom rather than making it your new default.

Watch the open issues, not just the announcement. Issue #31860 and issue #32806 are both still open as of this writing. Given that the effective window has already changed twice without a changelog entry, there’s no guarantee the manually configured 1M ceiling holds steady either.

If you’re comparing Codex against other agentic coding tools on context window alone, verify the number in practice, not the spec sheet. The same caution applies broadly — our coverage of Claude Sonnet’s 1M context window against GPT-5.3-Codex treated the advertised figure as the deciding factor at the time. This story is a reminder that an advertised context window and a shipped default aren’t always the same claim.

The Bigger Picture

This is a narrower story than the infrastructure and acquisition news we’ve covered this month, but it fits a pattern we keep running into: the gap between what a frontier AI product advertises and what it actually delivers by default, discovered not through an official disclosure but through users comparing notes in a GitHub issue thread. It happened with Anthropic’s Claude Code auto mode rollout, where the real story was in the data behind a default change, not the announcement itself. It’s happening here with a context window that quietly shrank twice before anyone at OpenAI explained how to get the number back.

Codex is competing directly against tools like Claude Code and Cursor on exactly this kind of spec — how much context an agent can hold before it starts forgetting. Our head-to-head coverage of GPT-5.3-Codex against Claude Sonnet 4.6 put Claude’s context handling ahead specifically because Sonnet’s 1M window, even in beta, behaved the way it was advertised. If Codex’s default context keeps drifting downward without disclosure, that’s a genuine competitive liability — one a config.toml workaround fixes for the users who find it, and does nothing for everyone else.

Our Take

We think Sottiaux did the right thing by posting the workaround publicly and naming the tradeoff in the same message — that’s more transparency than a silent cap deserves credit for, and it’s worth acknowledging. What we’d push back on is the framing that this was ever just “tuned for cost.” A default that drops from 353,400 to 258,400 effective tokens with no changelog entry, discovered by users because their sessions started compacting earlier than the week before, isn’t a tuning decision communicated to customers. It’s a tuning decision customers had to reverse-engineer from broken workflows and then compare notes about in a GitHub thread.

If you’re paying for Codex specifically because GPT-5.6 Sol’s 1.05M context window was part of the pitch, that number was never a promise about your default experience. It was a ceiling you had to know existed and manually ask for. Read the config.toml steps, decide if the cost tradeoff is worth it for your workload, and don’t take an advertised spec at face value again without checking what the product actually ships.

Frequently Asked Questions

How do I enable the 1M context window in Codex for GPT-5.6 Sol?

Edit ~/.codex/config.toml and add three lines before any [section] header: model = "gpt-5.6-sol", model_context_window = 1000000, and model_auto_compact_token_limit = 900000. Save the file, fully restart Codex, and start a new session — the override doesn’t apply to a session already in progress.

Does the 1M context override work with a ChatGPT Plus or Pro subscription?

It’s now confirmed to work for ChatGPT-billed accounts following Tibo Sottiaux’s August 17 post, though some reporting on subscription billing describes inconsistent behavior in practice. Previously, the override reportedly only reliably applied to API-key billing. If you’re on a subscription plan, verify your actual effective context during a session rather than assuming the config value is fully honored.

What is GPT-5.6 Sol’s actual context window in Codex by default?

Per GitHub issue #31860, Codex initially capped GPT-5.6 Sol at 372,000 raw tokens (353,400 effective). Issue #32806 documents a further, undocumented cut on July 13, 2026, down to 272,000 raw tokens (258,400 effective) — roughly a quarter of the model’s advertised 1.05 million-token spec.

Why does Codex cap the context window below the advertised spec?

OpenAI hasn’t published an explanation beyond Sottiaux’s comment that the default is “tuned optimally when it comes to performance and cost.” A larger context budget processes more tokens per turn, which costs more against usage-based billing or weekly quotas — but OpenAI hasn’t disclosed the cap publicly or explained the two documented reductions in the open GitHub issues.

Is there a downside to manually enabling the full 1M context window?

Yes. A bigger context budget means more tokens processed per turn, which burns through usage-based billing or a metered weekly quota faster than Codex’s tuned default. Reserve the override for sessions that genuinely need the extra headroom — long agentic runs across large codebases — rather than making it your default setting.

How does GPT-5.6 Sol’s context window compare to Claude or Grok?

GPT-5.6 Sol’s advertised 1.05M-token spec is larger than Grok 4.6’s 500K window and roughly comparable to Claude Fable 5 Max’s 1M window, per our three-way comparison. That comparison reflects the published spec, not Codex’s documented default cap — the two are not the same number in practice.

---

Last updated: August 19, 2026. Sources: Tibo Sottiaux on X · GitHub issue #31860 · GitHub issue #32806 · Codex official site · explainx.ai — Enable 1M-Token Context Window in Codex CLI.

Related reading: Grok 4.6 vs. GPT-5.6 Sol vs. Claude Fable 5 · OpenAI’s GPT-5.6-Cyber Crosses Its Own Risk Line · GPT-5.3-Codex vs. Claude Sonnet 4.6 · Claude Code Just Stopped Asking Permission