OpenAI launches GPT-6 Astra in what the company calls the “AGI era”

September 3, 2026 OpenAI officially unveiled a new AI model, the GPT-6 Astra. Rumors about this have been going on for months, and today they were confirmed: the company began rolling out the model to business customers, and in the coming days it will be available to ordinary users of ChatGPT. Collected the most important things in simple words.

## What happened

In a closed-door press briefing, OpenAI president and co-founder Greg Brockman concluded the presentation with the phrase: “Welcome to the AGI era” (AGI – artificial general intelligence, i.e. AI capable of performing almost any human intellectual work). This is an unusually bold statement, even by the standards of the industry, where announcements of new models are accompanied by loud promises.

The official name of the model is simply Astra, not "GPT-6" - OpenAI itself did not decide what to call it (there were variants like GPT-5.7). But in the API and in all the materials, the company ends up using a bundle of gpt-6-astra, so it's actually GPT-6.

## The main feature: AI, which itself controls the computer

If earlier neural networks mainly answered questions with text, then Astra can work ** directly in the program interfaces ** - the way a person does it: clicks, typing, switches between browser tabs.

According to OpenAI, the model can:

• fill out forms in the browser and update entries in CRM systems;
• work with tables and Power BI, analyze data in Python;
• collect sites and test them;
• work in engineering programs (for example, KiCad, FreeCAD);
• install and configure software;
• create documents, presentations and even simple 3D games.

At the presentation, a video was shown: an employee asks to turn a regular yellow circle into a rocket, and then into a full-fledged 3D game, and Astra does it in a couple of minutes. The idea is not to write a separate integration for each program – the model simply uses the same interface as the human.

## How smarter than previous models

XX

OpenAI calls Astra its largest training project ever, and claims that the leap in capabilities between the previous flagship model (GPT-5.6 Sol) and Astra was even greater than between Sol and its predecessors.

From the declared results on tests (benchmarks):

• 97.6% on FrontierMath Tier 4 v2 (complex math);
• 96% on GPQA Diamond (graduate level issues);
• 74.1% on DeepSWE v1.1 (complex engineering/software tasks);
• 98.6% on the ARC-AGI-3 test, which is designed to show the ability of AI to solve unfamiliar tasks, rather than simply repeat what it has learned.

It was the latter figure that caused controversy: some experts note that the high result can be explained not only by the model itself, but also by the “wiring” around it – tools, memory and the system in which it works. Brockman himself admitted that he did not consider 98.6 percent to be a mathematical proof of AGI, but rather an argument that a “qualitative shift” had already occurred.

## Astra is the first model with a “critical” level of risk in cybersecurity

A separate and rather alarming detail: OpenAI officially assigned Astra the highest level of risk in its assessment system – “critical” – in the field of cybersecurity. This means that the model, with the right tools, can independently find unknown vulnerabilities in well-protected systems and build chains to exploit them without constant human involvement.

Because of this, access to the most powerful “cyber” capabilities of the model is limited: first of all, they will be received by trusted organizations responsible for protecting critical infrastructure through a special program Daybreak Blue. For other users, these features will be available with additional restrictions and monitoring.

It is also known that shortly before the release, after the incident with the hacking of the Hugging Face (the Astra itself was not injured in it), OpenAI suspended some internal development for a couple of weeks to strengthen the protection of infrastructure – and only then resumed work on the model taking into account new control measures.

## Price and availability

The model has already started rolling out to corporate customers through the *Daybreak program. In the coming days, it will be available:

• chatGPT Plus, Pro, Business and Enterprise;
• via OpenAI API;
• through AWS Bedrock and Microsoft Azure cloud platforms.

API prices are much higher than previous models:

Режим	Вход (за 1 млн токенов)	Выход (за 1 млн токенов)
Standard	$10	$50
Fast (в 2,5 раза быстрее)	$20	$100

For comparison, the Claude Opus 5 costs $5/$25 and the GPT-5.6 Sol in standard mode costs $5/$30. That is, the Astra is the most expensive mass-available model on the market today.

At the same time, OpenAI states that comparing models by price per token no longer makes much sense: tokens of different models “weigh” differently, and it is more correct to compare the price for the task performed**. According to the company’s own data, on the DeepSWE v1.1 test, the best Astra configuration costs about 57% less per problem than the best configuration of GPT-5.6 Sol – even taking into account the higher price for the token.

## In Short: What You Need to Know

• OpenAI has released GPT-6 Astra – a model that is positioned as the beginning of the “age of AGI” and is able to work independently in the interfaces of programs, and not only respond with text.
• The official name in the API is gpt-6-astra; rollout starts for business customers and will reach ordinary ChatGPT subscribers within a few days.
• This is the first model that OpenAI has assigned to a “critical” level of cybersecurity risk, meaning that access to the most powerful features is limited.

The news is fresh - details (wider access, first-time user reviews, independent tests) are likely to be updated in the coming days. *

---

Sources: Bloomberg, CNBC, VentureBeat. *