GPT-5.6 for Everyone, Codex Inside ChatGPT: What Changed on Thursday and Why OpenAI Can Afford It

On Thursday, July 9, OpenAI released not just a new model. GPT-5.6 has arrived for all tiers, including the free one; ChatGPT Work was born, an assistant that works autonomously for hours on a project; Sites emerged, allowing users to build shareable sites and pages without knowing how to program; and the Codex app was integrated into the new ChatGPT app for computers, while Atlas, the browser launched just last October, is being retired. This is all in the official announcement.

It's worth looking at the whole picture because the model is just the most visible part, and the interesting question of the week is another: how does OpenAI manage to give away so much?

Three Models, Three Prices

The GPT-5.6 family has three versions: Sol, the most powerful; Terra, the middle one; and Luna, the most affordable. The names replace the old suffixes like mini and nano, and the number indicates the generation. For those using the models via API, i.e., within their own applications, prices are paid per token, which are the pieces of text in which the model breaks down everything it reads and writes: Sol costs $5 per million tokens read and $30 per million tokens written, Terra half of that, Luna even less, per the official pricing, found at the bottom of the announcement page (for readers looking for specifications: context of one million tokens, maximum output of 128,000, knowledge updated to February 16, 2026). The substance: the best model costs as much as the previous one, the middle one promises similar performance to the old top model at half the price, and the small one opens a new economic segment for simple and repetitive tasks.

Note: In the first hours after launch, the GPT-5.6 selector wasn’t available on my account, and it only appeared after I manually downloaded the app update from the OpenAI website on Friday. The update apparently isn’t automatic. So if you don’t see the model, try re-downloading the app from the announcement page before taking it out on the rollout. The rollout itself is declared global and without exclusions for Europe, unlike Grok 4.5, which SpaceXAI will bring to the EU only in mid-July, as reported to Reuters.

The Model Now Decides How Much to Think

The key technical novelty is one: Sol can "think" for varying lengths of time before responding; there is a maximum reasoning level for difficult problems and an ultra mode where multiple copies of the model work in parallel on the same task. Thinking costs, as reasoning is also paid for in tokens: Simon Willison calculated, using his now-famous pelican on a bicycle test, that the exact same request can cost anywhere from less than a cent to almost half a dollar, depending on the model chosen and how much thinking time is allowed. There’s also a novelty in caching, the memory that prevents re-reading (and re-paying) for parts of the conversation that are always the same: it’s now more predictable, and for developers, it’s the simplest way to save.

ChatGPT Work, Sites, and the End of Atlas

However, the biggest product novelty is not the model; it’s what has been built around it. ChatGPT Work is an assistant to whom you give a goal rather than a command: it breaks it down into steps, works for hours, and ultimately delivers spreadsheets, documents, presentations, or small applications, pulling from the files and apps you’ve connected.

Does this remind you of anything? It does to me: it’s the same process employed by Anthropic’s Fable 5: you give it a task, and it finalizes it for you without much question, delivering the final result after some reasoning.

Under ChatGPT Work lies the same technology as Codex, the programming tool that OpenAI claims is used every week by over 5 million people. The initial examples mentioned in the announcement are customer testimonials, and while they should be taken as such, they convey the idea: at NVIDIA, a job in Excel that consumed nearly half the preparation time of an event is now handled by the agent, and at Virgin Atlantic, an analysis that used to take weeks is down to hours. Using the agent consumes the quota included in the plan, so it’s not unlimited.

Surrounding the agent, OpenAI has reorganized things. The Sites transform a project into a shareable site or small web app accessible via a link. The computer app becomes a single one, containing chat, agent, and coding, available on every plan including the free one. And Atlas, the autonomous browser presented less than a year ago, is closing: its functions are being absorbed into the app and the Chrome extension. This too says something: even OpenAI doesn’t get everything right on the first try.

The Scores, with a Pinch of Salt

And how good is the model? Labs compare themselves with benchmarks, score competitions on standard tasks. OpenAI claims the record on a long-duration professional job contest, scoring 13.1 points above Claude Fable 5. However, on the most cited programming competition, Fable 5 remains significantly ahead, scoring 80% against 64.6%, and on the eve of the launch, OpenAI released a study stating that about 30% of the exercises from that contest would be defective. They may have a point; these tests have known issues, but the lesson for readers is simpler: every lab presents the contest in which it wins, so the scores on launch day should be read with detachment. The only independent impression so far comes from Willison, who previewed Sol: very competent, but for difficult programming tasks, not better than Fable. If you use these models for work, the test that counts is yours: try them on your real tasks before making a change.

How Does OpenAI Afford to Give So Much? With Gigawatts

The answer to the initial question lies right here: OpenAI claims to have exceeded 10 gigawatts of guaranteed computing infrastructure, a goal set for 2029 that has been reached three years ahead of schedule, and the generosity of the plans is the strategy of those with abundant capacity compared to competitors, using it to occupy space.

A Line on the Context

One last thing to know: before the public release, GPT-5.6 underwent two weeks of private preview with about twenty organizations agreed upon with the U.S. government, and the final green light came from government leadership, as reported by Bloomberg. This is the same path already seen with Anthropic’s models, which remained off for 18 days in June by order of the Department of Commerce, as detailed on Anthropic’s blog. The most powerful models now arrive on the market with an additional institutional step, and it’s worth knowing this even just to better read the upcoming weeks.

In Practice

If you’re already using GPT-5.5 for work, the candidate for a change is Terra, to be validated on your real tasks. If you need low-cost volume, try Luna. The new computer app deserves a look just out of curiosity because it shows where the whole category is heading. Now that the selector has also appeared for me, in the coming days I will compare it with Fable 5 on the weekly tasks; if you’ve already done so, please share in the comments what worked and what didn’t.