GPT-6.1 Sol's Default Context Just Went From 272K to 500K in Codex: What It Changes, and What It Doesn't

GPT-6.1 Sol — the mid-tier model released on 2026-09-29 at $2.00 per million input tokens and $10.00 per million output — had its default context length in Codex and ChatGPT Work raised from 272,000 tokens to 500,000 on 2026-10-12, an increase the vendor puts at roughly 80 percent. Thibault Sottiaux, the company's head of product and platform, announced it as Day 7 of the company's 28-day shipping sprint, described 500,000 as a thoroughly tested "sweet spot," and said the default would go higher again in future. The maximum is untouched and so is GPT-6 Sol, the generation 6.1 replaced, whose published ceilings this one inherited unchanged.

That is the whole announcement, and it is easy to read past the two words doing the work. Default, not maximum. Codex and ChatGPT Work, not the API. The model has been able to accept far more than 500,000 tokens since it shipped — the number that moved is the one you get when you ask for nothing.

Everything below is read from the vendor's own model and pricing documentation and from its announcement, both on 2026-10-12, or is arithmetic on the rates it publishes. Where a figure is the vendor's own and nobody outside has reproduced it, the sentence carrying it says so.

The delta, itemised

Six things are worth separating, because three of them moved and three did not.

• Default context in Codex and ChatGPT Work — 272,000 tokens before 2026-10-12, 500,000 after. This is the number that applies when a request does not specify one.

• Maximum context — unchanged. the vendor's model page for GPT-6.1 Sol lists a 1,050,000-token window, a maximum input of 922,000 tokens and a 128,000-token output cap. Those figures are identical on the page for GPT-6 Sol.

• Manual override — unchanged. Users can still set the working context higher or lower than the default, up to the model's own ceiling, which is why none of this removes an option that already existed.

• The subscription long-context surcharge — removed for GPT-6.1 Sol, explicitly "under ChatGPT subscription quotas." This is the clause with money attached, and it is limited to usage drawn from a ChatGPT plan.

• API defaults and API pricing — unchanged. The API belongs to you, not to a client's default table, and the long-prompt multiplier on the pricing page still reads as it did before the announcement.

• The 1M ceiling — no change, but worth stating because a 500,000-token default spends 54 percent of the 922,000-token maximum input allowance. "It fits in 1M" and "it fits under the default" are now two different questions with a 422,000-token gap between them.

The one number in the list that a reader can act on today, therefore, is not 500,000. It is the surcharge waiver — and only if their usage is billed against a subscription rather than an API key.

The surcharge is the part that changes a bill

Run the arithmetic on the vendor's published API rates and the shape of the problem the waiver addresses is obvious. GPT-6.1 Sol's listed rates are $2.00 per million input tokens, $0.10 per million cached input tokens and $10.00 per million output tokens. Prompts above 272,000 input tokens are priced at 2x input and cache rates and 1.5x output for the full request — not just for the tokens past the line.

• 270,000 uncached input tokens, 10,000 output — input 270,000 at $2.00 is $0.540; output 10,000 at $10.00 is $0.100. Total, $0.640.

• 300,000 uncached input tokens, 10,000 output — the request is over the line, so every rate moves: input 300,000 at $4.00 is $1.200; output 10,000 at $15.00 is $0.150. Total, $1.350.

Eleven percent more input tokens, 2.1 times the bill. The step is what makes a larger default a genuine question rather than a free upgrade, and it is the step the vendor just took off the subscription side for this model. On an API key it is still there, on exactly the same threshold, unchanged by anything announced on 2026-10-12.

If you call the model through OrcaRouter, that boundary is the vendor's and travels with it: we serve openai/gpt-6.1-sol at the vendor's list rates with 0% markup, so the $2.00 / $0.10 / $10.00 card and the 272,000-token multiplier above it are the same numbers you would see on the vendor's own page, and a change on either side lands on our side the same day rather than at a repricing window. What we do not do is change the vendor's tier. If a request crosses 272,000 input tokens on an API key, it is repriced, whoever routed it.

Where the new default actually lives

The announcement names two surfaces, and the distinction matters more than it sounds. Codex and ChatGPT Work are the agentic products; GPT-6.1 Sol is not in standard ChatGPT chat. So the 500,000-token default reaches the people running long coding loops and document-heavy work sessions, and reaches nobody calling the model from a script.

There is a second layer under that. A client connecting to the model over ChatGPT subscription credentials is not the same shape as a client connecting with an API key — it carries a budget table that caps eligible models so that usage drawn from a subscription matches what Codex itself allows. One open-source coding client, OpenCode, merged a change to its ChatGPT-OAuth connection the same day, raising its entry for gpt-6.1-sol from 272,000 to 500,000 input tokens against a 628,000-token context ceiling, "preserving the existing 128k output headroom." Its reviewer's note is the honest summary of the whole announcement: only the default limits for this model on that connection changed, the public API is unchanged, and per-model config overrides still apply.

That is a third-party client's reading of how the caps are applied, not a vendor document, and it is worth labelling as such. But it matches the announcement's own framing: the vendor raised a default it controls, on the surfaces where it sets the default, and left the API where the caller decides.

Why the vendor moved the default instead of the ceiling

The stated reason is compaction. When a long-running Codex session approaches its working context, the harness condenses earlier turns — summarising and discarding detail — so the session can continue. Compaction is not data loss in the crash sense; it is a deliberate trade of fidelity for headroom, and it fires more often the smaller the default is. Raising the default to 500,000 tokens means the trigger arrives later on exactly the tasks that need retention most: multi-file refactors, long debugging arcs, work carrying a large retrieved corpus across many turns.

This is also why the announcement is about a default rather than a limit. A maximum only helps a caller who knows the limit and sets it; a default helps every caller, including the ones who never configured anything. For a product whose users are largely interactive, that is the higher-leverage knob. It is the vendor's own rationale and it is reasonable on its face — the measurable half of the claim is that fewer compactions occur, which is a harness behaviour rather than a model capability, and the model itself has not changed. Same snapshot, same rates, same 128,000-token output cap, same knowledge cutoff of 2026-04-30.

The case against, stated as opinion

Practitioner reaction was not uniformly positive, and the objections are worth carrying because they are testable claims, not vibes. One developer publicly called the change an anti-feature on two grounds: subscription quotas will be consumed faster if sessions routinely run longer before compaction, and needle-in-a-haystack retrieval quality is unlikely to be as good at 500,000 tokens as at 272,000, so the extra room may not be usable room. Others made the related point that a larger window is a capacity, not a recommendation — filling it with loosely relevant files dilutes attention and slows responses.

These are community positions, not vendor guidance, and the vendor's counter is implicit in the word "sweet spot": the value was tested and chosen for experience as well as capability. But the two concerns point at different things and should be separated. The quota point is an accounting question — if your plan meters tokens, a longer default means more tokens before the safety net of compaction kicks in. The quality point is an evaluation question, and it is answerable the only way these questions ever are: run your own long-context task at both sizes and compare, rather than taking either the announcement's framing or a reply thread's.

What to check before you rely on the new number

• Which surface you are on. Subscription-billed Codex or ChatGPT Work gets 500,000 as the default and no long-context surcharge. An API key gets neither — no changed default, and the 272,000-token multiplier still applies.

• Whether your harness sends an explicit limit. A client that hardcodes a context value overrides the new default entirely. Check the config before assuming the bump reached you, because the fix may be a line in your own settings rather than anything the vendor shipped.

• How the caps compose. A third-party ChatGPT-OAuth connection may present its own table — the OpenCode case above pairs a 500,000-token input allowance with a 628,000-token context ceiling and leaves 128,000 tokens of output headroom. If your client is in that category, verify its number rather than the announcement's.

• Whether your bill is a subscription or a meter. The surcharge waiver is a subscription-side change. The API tier is untouched, so a cost model built on the API has nothing new in it.

• Whether you should fill the window. The default is what you get; it is not a target. Cached input at $0.10 per million against $2.00 uncached is the lever that actually lowers a long-context bill, and it rewards byte-stable prefixes rather than bigger ones.

The open question

The vendor has said the default will rise again, which makes 500,000 a waypoint rather than a destination — and makes the interesting question not whether the number grows but whether the two constraints around it move with it. The 272,000-token repricing threshold sits below the new default now, which means a subscription-billed session that runs to its default on the API's rules would be a long-context request; the waiver is what stops that from mattering on the subscription side, and nothing announced on 2026-10-12 suggests the API threshold follows the default upward. Watch that threshold, not the window.

For a team already building on GPT-6.1 Sol, the practical reading is narrow and clear. If you are in Codex or ChatGPT Work on a subscription, you get more room before compaction and one fewer line on the bill, starting now. If you are calling the API, nothing changed, and the number worth watching is still 272,000. Our model page for openai/gpt-6.1-sol — 1,050,000-token context, 128,000-token output, $2.00 in and $10.00 out — is where that stays priced.