GPT-6.1 Sol: 1M Token Context in Codex, No Surcharge

What happened

OpenAI has roughly doubled the GPT-6.1 Sol context window in Codex and ChatGPT Work, and quietly killed the long-context surcharge that used to make marathon sessions expensive. If you've ever watched an AI coding assistant forget a decision you made twenty minutes earlier, this update is aimed straight at you.

Here's the short version. The standard context window — the amount of text the model can keep "in mind" at once — grew from 272K to 500K tokens. That's an 84% jump, not a rounding error.

A token, by the way, isn't a word. It's a chunk of text, usually a few characters or part of a word. Think of it like a moving company counting boxes instead of listing every item inside. So a 500K-token context window is a genuinely enormous pile of material — several novels' worth held in the model's working memory at once, rather than the few pages you'd normally paste in by hand.

And if 500K isn't enough, you can stretch it further, up to 1 million tokens. The real ceiling is a hair higher than the original announcement suggested: the model card lists 1.05 million tokens as the maximum, per the GPT-6.1 Sol model docs.

The pricing change is the bigger story

The second half of this update matters more than the first. OpenAI removed the long-context surcharge for subscribers using the quota included in their plan. Before, opening the context window all the way could burn through your limits noticeably faster. Now there's no separate surcharge stacked on top — your normal subscription limits still apply, but you're not getting penalized twice for the same long conversation.

The catch is that this doesn't carry over to the API. If you pay per token through the API, long-context pricing is still tiered, and requests above 272K tokens may still cost more per token. The thresholds are laid out in this breakdown of long-context pricing cliffs.

Why 272K keeps showing up

That's the old ceiling, and it's also the line where API pricing still shifts. In practice, 272K tokens was plenty for a solid day of coding but nowhere near enough for an entire codebase, a full set of contracts, or a novel-length manuscript. That's the gap this update closes.

What it means for you

You don't need to be a developer to care. Every AI chat hits the same wall eventually: the model starts summarizing, forgetting, or contradicting itself because its memory ran out of room. This update pushes that wall much further away. Here's what that looks like in practice.

At home

Planning a two-week trip with flights, hotels, visas, and a dozen "actually, let's skip that museum" revisions used to mean starting a fresh chat every few days. With a 500K window, you can keep one thread alive and have the model remember the dietary restriction you mentioned on day one.

At work

Drop in a full repository, a stack of meeting transcripts, and last quarter's reports, then ask questions across all of it. Developers in Codex get fewer "let me summarize what we've discussed so far" interruptions — and if you've used these tools, that's exactly where the real time savings hide. This write-up on the Codex context window walks through what it looks like day to day.

In business

Support teams can paste in a knowledge base plus a customer's entire history and get an answer that respects both. Marketers can hand over a brand book, three years of past campaigns, and a product spec, then ask for copy that doesn't drift off-brand by paragraph four. If you want to see how a longer context changes your own workflow without spending a cent, the free AI tools at mykreatool.com are a good place to start poking around.

For studying

One textbook chapter is nothing anymore. A whole semester of lecture notes, past papers, and your own summaries is now fair game in a single conversation. You can ask "what did my professor emphasize in week 6 that connects to week 11?" and get an answer drawn from both.

For creative work

Novelists and world-builders are the quiet winners of every context bump. Keep your character bible, timeline, and your entire draft in one chat, and the model stops giving your protagonist two different childhoods. iWeaver's breakdown covers more of the creative angles.

For income

If you freelance, longer context means fewer hours spent re-explaining a client's project every Monday morning. Fewer resets, faster turnarounds, less cleanup after an AI that forgot the brief. That's billable time either recovered or lost — your call which.

How to try it right now

The free option is doing nothing at all. If you already have GPT-6.1 Sol in Codex or ChatGPT Work, the 500K context window is the default, and there's no long-context surcharge nibbling at your plan's quota. Just start a session and use it.

If you want the full 1M context, you'll need the local Codex CLI or IDE — about two minutes of work.

Step 1. Open your Codex config file:

• macOS and Linux: `~/.codex/config.toml`

• Windows: `%USERPROFILE%\.codex\config.toml`

Step 2. Add or edit these three lines:

```toml

model = "gpt-6.1-sol"

model_context_window = 1000000

model_auto_compact_token_limit = 900000

```

Step 3. Save the file and restart Codex so it picks up the change.

The last line is the one people skip. It sets the point at which the tool automatically compacts — meaning summarizes and trims — the conversation, at 900K tokens. Set it too close to the ceiling and you'll slam into the limit mid-task; set it too low and you're back to losing details for no good reason.

Full details live in the Codex config reference.

Upsides and what changes

The obvious win is that long projects stop falling apart halfway through. You'll re-explain less, re-paste less, and spend less time correcting a model that lost the thread three hours ago.

For subscribers, the pricing change is the bigger deal. Long-context work now runs on your existing plan quota, which makes costs far more predictable. You can plan a heavy research day without wondering whether the last hour just ate a week of usage.

For API users, the shift is smaller but real. You get the larger window, but tiered pricing above 272K tokens means you should still keep an eye on per-token cost before you shovel in a million tokens.

Limitations

Let's be straight about the fine print. A million tokens is not infinite memory — it's a lot, but the model can still lose track of details buried in the middle of a giant prompt, and quality doesn't scale perfectly with raw size. The hard ceiling is 1.05 million tokens, and only GPT-6.1 Sol gets these numbers. Subscription limits still exist; removing the surcharge didn't remove the cap. API and token-based billing follow separate rules, with higher rates for requests above 272K tokens, so anyone paying per token should run the math first. Bigger contexts aren't free in the physical sense either — longer prompts mean slower responses and, on the API, higher cost per call. And you'll need a plan that includes GPT-6.1 Sol in Codex or Work before any of this applies to you.

Conclusion

OpenAI just gave its coding and work assistant roughly twice as much room to think, and stopped charging subscribers a premium for the privilege. For anyone who has spent a Monday morning re-pasting a project brief, that's a genuinely useful upgrade rather than a spec-sheet flex.

One action for today: open `~/.codex/config.toml` (or `%USERPROFILE%\.codex\config.toml` on Windows), decide whether the default 500K or the full 1M suits your work, and set `model_auto_compact_token_limit` so long sessions compact before they hit the wall.