GPT-6.1 Sol launches: stronger coding and agent work at the same standard price

Better AI gets a lot more interesting when you can afford to use it for the whole job.

That is the appeal of GPT-6.1 Sol. OpenAI introduced it on 29 September 2026 as an upgrade for coding, computer use and professional work, with performance approaching Astra on selected tests. The standard input and output prices stay the same as GPT-6 Sol. OpenAI’s launch announcement puts them at one-fifth of Astra’s.

The interesting question is how much work this moves out of the premium-model category. If Sol can handle more of a project without needing the expensive model to step in, the benefit goes well beyond a cheaper answer.

This is the next chapter after the original GPT-6 Sol launch. It sharpens the same proposition: capable AI for the work you want to do regularly, at a price that makes regular use plausible.

Key points

• The upgrade is about capability at the same base price. OpenAI reports stronger coding, computer use and document work than GPT-6 Sol.
• Standard API input and output remain US$2 and US$10 per million tokens. Cached input reads drop to US$0.10.
• Look in Work or Codex. Ordinary Chat was explicitly outside the launch; the subscription rollout starts with Pro.
• Near-Astra is a claim about selected evaluations. It does not mean the two models are interchangeable on every job.
• Sol Ultrafast was announced for later. The available Astra speed tier has separate access and pricing.

Why the price matters more than another model number

An AI task rarely ends with the first answer. You ask for a change, spot something it missed, add another document or realise the original question was slightly wrong. Each turn adds to the work the model has to do.

Imagine building a small app. The first version needs a form, the form needs validation, the validation reveals an awkward assumption, and fixing that assumption changes another part of the app. A useful assistant has to follow the project through those turns. A cheap first draft that cannot carry the next step is only a small part of the solution.

That is why I find Sol’s positioning more interesting than the familiar race to declare the most powerful model. A capable middle tier can change how freely people use AI. You can give it the surrounding job, rather than rationing the good model to a few carefully chosen questions.

There is a consumer benefit hidden inside the developer price story too. Developers who can run useful features at lower cost have more room to offer them at a reasonable price. That will not automatically flow through to every app subscription, but it makes more ideas economically possible.

What GPT-6.1 Sol costs

Here are the standard short-context API rates. Tokens are the pieces of text the model reads and produces; these prices are for developers paying for model usage, rather than the monthly price of a ChatGPT plan.

The straightforward comparison is the input and output rate: Sol costs one-fifth as much for the same quantities. Cached reads have a different discount. Caching lets a model reuse eligible material, such as stable project instructions, at a reduced read price.

The cheap cached rate deserves a precise reading. It applies to reading cached input, rather than every piece of material you send. OpenAI’s caching guide explains the separate charge for cache writes.

There is also a size threshold. The model reference says prompts above 272,000 input tokens attract higher rates for the whole request: twice the input and cache rates, and 1.5 times the output rate. Other processing options have separate prices. These are US-dollar figures; they should not be read as Australian-dollar charges.

For a simple hypothetical example, suppose a request uses 10,000 uncached input tokens and 2,000 billed output tokens at the standard short-context rates. The model charge would be four US cents on Sol and twenty US cents on Astra. That excludes tools, retries, taxes and other charges, and assumes both models use exactly the same quantities. It is arithmetic, rather than a measured result.

Four cents sounds trivial. Repeated across a busy product, the difference becomes meaningful. The remaining question is whether the cheaper model produces work good enough to keep.

What the stronger model could change

OpenAI’s reported gains cover coding across real projects, working with software interfaces and answering questions from complex documents. Those are useful areas to improve because the challenge often lies in connecting several details correctly.

Consider a report with a chart on one page, a definition on another and an exception buried in the notes. Answering a question about it requires more than finding a sentence containing the right words. The model has to connect the information and preserve the exception.

Or think about a bug whose visible symptom is on a website, while the cause lives in the data arriving from somewhere else. A model that understands the whole path is more useful than one that patches the first thing it sees.

The launch results make Sol a more credible candidate for those jobs. They do not settle the result in your particular project, and Astra still has a role in demanding work. “Near” is doing some work in the near-Astra description.

I’d read this as an expansion of the jobs worth giving Sol, rather than a reason to retire every more capable model. The most awkward cases tend to be the ones where the brief is incomplete or the answer requires a judgement nobody has made yet.

Where you can actually use it

The rollout has an easily missed detail: Work, Codex and ordinary Chat are different places. The launch includes the first two and the API.

OpenAI’s current release notes describe the Pro-first rollout. The API availability and ordinary-Chat exclusion come from the launch announcement. An eligible subscription does not guarantee that the rollout has reached your account.

If you mainly use ChatGPT for quick questions, this release may have less immediate effect on your day than the headlines suggest. It becomes more relevant when you use Work to produce something substantial, Codex to work on software or an app built on the API.

Big documents fit, but the job still needs a clear brief

In the API, Sol’s documented capacity is a 1,050,000-token context window and up to 128,000 output tokens. Context is the material it can consider; output is what it can produce. It accepts text and images and produces text.

That is room for substantial source material. It also creates a temptation to send everything and hope the model finds the point.

I would still tell it which decision you need, which material matters and what a useful result should contain. A large filing cabinet is handy. Emptying the filing cabinet onto someone’s desk is a different strategy.

For example, an instruction to compare three proposals should identify the requirements that decide the choice. Otherwise the model can give you a beautifully arranged comparison of features you do not care about.

The capacity figures make bigger jobs possible. A clear brief makes them worth doing.

What is still on the way

At the 29 September launch, OpenAI said Sol Ultrafast would follow in the coming days. As checked on 3 October, its API and Work/Codex guides describe Astra Ultrafast; they do not establish a Sol rollout date.

Keep the two decisions separate. Sol’s standard price makes capable work cheaper. An Ultrafast tier is about paying for less waiting. Having access to the model does not automatically give you its announced premium-speed option.

Developers moving an existing application also have compatibility work to do. OpenAI’s migration guidance covers changed reasoning settings and the requirement to use Responses for tools. A production application needs more care than replacing a model name in a menu.

The upgrade that could become the everyday choice

The real opportunity for GPT-6.1 Sol is to be capable enough that you stop thinking about whether a routine project deserves the expensive model.

That does not require Sol to win every comparison. It requires it to handle a broad enough range of useful work at a price that lets you keep going.

I’d start with something you already know how to judge: a report you understand, a change with a clear expected result or a set of documents with questions you can verify. If Sol can carry more of that job without constant correction, the upgrade has earned its place.

The model number will become familiar quickly. The more lasting change could be how much work people feel comfortable handing it.