GPT-6 Astra: what OpenAI's new model actually changes

On 3 September 2026 OpenAI released GPT-6 Astra, its new flagship model and the successor to GPT-5.6 Sol. The company calls it its most intelligent and aligned model so far, with the largest gains in computer use, browsing, software engineering, cybersecurity, science and professional work.

A note on the name first, because the search traffic is split three ways. People are looking for "ChatGPT 6", "ChatGPT Astra" and "Astra 6", but the official name is GPT-6 Astra and the API identifier is gpt-6-astra.

The interesting question is not whether the benchmark numbers are large. They are. It is what changes in the systems the rest of us build on top.

From talking to an assistant to delegating to one

Each generation of ChatGPT got better at understanding an instruction and returning a good answer to it. Astra is tuned for something different: finishing a job that takes many steps, using whatever computers, browsers, files, applications and tools it has been given access to.

The shift is from "tell me how to do this" to "do this". Instead of explaining how to research a market, pull the numbers together and build the deck, the model is meant to work through those stages and hand back the finished thing.

That matters most for the people whose day is spent moving information between documents, spreadsheets, websites and internal systems. That work is not difficult. It is just long, and it has never automated well, because every step needs a small judgement a script cannot make.

Computer use is the headline capability

OpenAI positions Astra as a new frontier in computer and browser use: navigating unfamiliar interfaces and carrying out multi-step workflows inside them.

That changes the role an assistant plays. It stops being the thing that tells you to open the site, find the report, download it and analyse it, and starts being the thing that does those four steps, within whatever permissions you granted it.

The obvious candidates inside a business:

• Research and data collection across sources that have no API.
• Back-office admin: form filling, reconciliation, moving records between systems that will never be integrated properly.
• Document processing and the reporting built on top of it.
• Software testing against a real interface rather than a mock.
• Support workflows where the answer means looking something up in three different tools.

The useful way to frame it: computer use turns a model into an execution layer, not just a conversational one. That is a genuinely different thing to secure, log and review.

What it means for software work

OpenAI calls out software engineering as one of the areas where Astra is state of the art, and this is where the difference between generations is easiest to feel.

Real development work is rarely a snippet. It is reading an unfamiliar codebase, finding where a bug actually originates, changing several files consistently, running the tests, reading the failures and adjusting. The hard part has always been holding the whole task in view across those steps without drifting off the original requirement.

Astra is built for those longer loops, and OpenAI specifically claims it is better at holding on to the original objective when new instructions arrive mid-task. If that survives contact with a large codebase, it is worth more in practice than another few points on a coding benchmark.

Professional work, to your template

The other focus is professional output: documents, presentations, spreadsheets and analyses that follow an existing template and an existing house style.

That is a harder problem than it sounds, because a usable business document has to do six things at once:

1. 1Understand what the document is actually for.
2. 2Find the information that bears on it.
3. 3Leave out everything that does not.
4. 4Organise what remains so a reader can follow it.
5. 5Obey the required format.
6. 6Produce something usable as it stands, not something to be edited into shape.

Give the model the financial data, the internal notes and last quarter's deck, and the goal is a business review that fits the way the company already works. Generic quality is not the target. Fitting the existing workflow is.

The benchmark numbers, and how to read them

The reported results are striking: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on the ExploitBench cybersecurity benchmark against 78.5% for GPT-5.6 Sol.

Read the ARC-AGI-3 figure carefully, though. That score depends on a stateful evaluation harness and a very expensive run. Independent stateless testing by the ARC Prize team lands far lower, across a wide band depending on the reasoning settings used. The number is real, but it describes a particular setup rather than a property of the model on its own.

What the scores do tell us is where the frontier is being pushed: planning, tool use and operating inside unfamiliar environments, rather than recall and fluency. That is the direction worth building for.

Capability brings a real security problem

Astra is the first OpenAI model to reach the Critical cybersecurity level under the company's Preparedness Framework. In plain terms: given tools and access, it can find previously unknown vulnerabilities and work out how to exploit protected systems without being walked through it step by step. Two zero-days were reportedly found during evaluation.

Hence the extra monitoring, isolation and access controls, and a gated rollout by trust level rather than a single switch flipped for everyone.

OpenAI also reports fewer factual errors than GPT-5.6 Sol, while its own system card is explicit that the evaluations do not establish reliability in every environment, and documents a measurable decline in how monitorable the model's chain of thought is.

The practical conclusion is the one that applies to every model I have put into production: a more capable system is not an unsupervised one. Anything financial, legal, medical or security-critical still needs a human on the decision, and the more of the work the model does, the more deliberate you have to be about where that review sits.

What changes for everyday users

If you use ChatGPT to explain a concept, draft an email, summarise an article or think through an idea, Astra will feel familiar. Those tasks were already handled well, and there is a ceiling on how much better a good answer to a small question can get.

The difference shows up when the objective is larger than one reply. The old pattern was ask, read, copy something out, ask again, then do the next step yourself. The new one is to state the objective and let the model research, analyse, use its tools and come back with a result to review.

The change that matters is not better answers. It is that the unit of work moves from the question to the task.

Availability and pricing

Astra began rolling out on 3 September 2026, first to a limited set of organisations and then more widely, across the Plus, Pro, Business and Enterprise plans, with what you get depending on the plan.

For developers it is available through the API as gpt-6-astra, with a 1M-token context window, at $10 per million input tokens and $50 per million output tokens. That is roughly 2.5 times the price of GPT-5.6 Sol, which is the number to hold on to when you size a workload.

Worth saying plainly: at that price, routing every request to the flagship is a design decision, not a default. Most production systems I build still send the bulk of traffic to a smaller, cheaper model and keep the frontier model for the steps that genuinely need it.

The bigger picture

The industry is moving from AI that generates content towards AI that understands an objective, uses tools, operates software and completes work. Astra is a marker along that road rather than a turn off it.

It does not remove the need for people. It moves where their attention goes: defining the objective, supplying the context, setting the boundaries of what the system may touch, and reviewing the decisions that matter. Those are the parts that are hard to specify and expensive to get wrong.

For anyone building on this, the engineering questions are the familiar ones. What is this agent allowed to do, how do we find out when it was wrong, and what happens next when it is. The model got better. That checklist did not get shorter.

Frequently asked questions

What is GPT-6 Astra?

OpenAI's flagship model, released on 3 September 2026, built for advanced reasoning, computer and browser use, software engineering, science, cybersecurity and professional work.

Is Astra 6 the official name?

No. "Astra 6" and "ChatGPT Astra" are common search terms, but the official name is GPT-6 Astra.

Is GPT-6 Astra better than GPT-5.6 Sol?

OpenAI positions it as the more capable flagship, with gains in computer use, reasoning, software engineering and professional work. It also costs about 2.5 times as much per token, so better does not automatically mean right for a given workload.

Can GPT-6 Astra use a computer?

Yes, and it is the headline capability. It can operate a computer environment and carry out multi-step workflows wherever the tools and permissions have been granted to it.

Can developers use it through the API?

Yes, under the model identifier gpt-6-astra, with a 1M-token context window, at $10 per million input tokens and $50 per million output tokens.

Is GPT-6 Astra reliable enough to run unsupervised?

No model is. OpenAI reports meaningful gains in factual accuracy and alignment, and its own safety documentation sets out the remaining limits and the uncertainty in those evaluations. Keep a human on high-impact decisions and keep the agent's permissions narrow.