GPT-6 Astra – What It Is And What It Means For Developers

September 03, 2026, OpenAI announced their latest model, named GPT-6 Astra or GPT AGI 6. According OpenAI, Astra is state-of-the-art on science, professional work and especially on computer us. With high scores on FrontierMath Tier 4, ARC-AGI-3 and ExploitBench, which they say it has 100% score on, it already brings solutions to standing open problems in mathematics.

This is not a full review. The model is not yet fully rolled out to all and I don’t access to it yet. This is more of a summary of what is known so far and what I think is interesting about it.

---

As always every new model announcement comes out with benchmark numbers. As we can see on the images below, quite amazing..but aren’t they all.

What caught my eye when I first saw the presentation video is that it can sit your computer and do all the work and all you have to it just ask it. You give it a task like “fill this form” or create me a “rocket image” which then turns into a playable game and instead of telling you how to do ti, it does it for you.

That is something different from what we had until now. I mean it gives us an idea of which way things are going and I am not sure if this is something good or bad in the long run.

What is GPT 6 Astra?

• Its a model name GPT 6 and that just means it’s a model in a league of its own. We thought there will be GPT 5.7, but they gave us something way out of the GPT 5 league.
• It was announced on September 03, 2026 and released as a limited preview with access given to a small number of organizations.
• Context window of about 1 million tokens and limited output to 128K tokens. Knowledge cutoff date is April 30, 2026.
• API pricing is $10 per million input tokens and $50 per million output tokens.
• It is available through OpenAI API, Azure and AWS Bedrock.

Computer Use

Now this is where Astra differs from the others. The model works like a person would. With no need for a dedicated API integration for every application, it looks at the screen, understands what is there, clicks, writes, navigates between websites, browsers, desktop applications, basically through anything you need and all you need is use your voice only.

For me as a developer this is the interesting part and kinda little scary. I can see it already replacing Alexa as home assistant. Why not have an assistant that can control your smart home AND control your computer and your internet appearance. All that in one smart personal assistant. The work we developers do is moving data from one system to another, filling forms, creating those forms, clicking around dashboards and for lot of this we already had scripts we write to make our life easier even before AI became a thing, but now with this model we don’t even need to write and ask things to do, just need to say it. Scary where this is going indeed, but we are nothing if not adaptable in new environments.

Cybersecurity and safety

Astra is the first OpenAI model that reached what is called “Critical” level for cyber capabilities in the Preparedness Framework. In other words, with the right tools and access, it can find security flaws that were previously unknown in well-protected systems and write exploits for them with a person guiding each step.

And this is what put a stop on the release and got it delayed. I think we all know the “Hugging Face incident“. They also said the strongest cyber features will stay restricted and that they have a system that monitors the model in the background. I don’t know how much of this is real, but it is the first time I see a company slow down a launch for this reason and say it publicly.

What is new for developers in the API

A few things from the documentation:

Async tool calling – the model can keep reasoning or call other tools while your application is still running a previous tool. You mark the tool as async and return the result later with the original call id.

• Mid-turn steering – you can send a correction or a new requirement while the model is working, and it continues with the updated instructions instead of starting over.
• Reasoning effort – goes from low to max. There is no “none” option, so reasoning is always on.
• Everything from GPT 5.6 still works – structured outputs, streaming, prompt caching, computer use, multi-agent orchestration.

Can I tell it to close the blinds?

Since I was curious about this one. There is no built-in smart home support in Astra. It doesn’t know what is on your local network. But if you already have a hub like Home Assistant that exposes your devices through an API, you can give the model a set of functions – set_blinds(), start_coffee(), set_temperature() – and let it decide which one to call from your voice command. This is the same tool-calling pattern people have used with previous models, only now the model is better at chaining multiple steps.

Honestly, for “turn off the lights” Alexa or Google Home will always be faster. Where a model like this makes sense is for “fuzzy” requests that need some thinking, not for simple on/off commands.

Conclusion

I think Astra is worth paying attention to, but not for the AGI headlines. It is worth because it is the first model where the main product is “it uses the computer for you” and not “it chats with you”. If it works half as good as the demo, it will change how we think about automation and integrations.

The public rollout is happening now, so in a week or two there will be real user reports instead of staged videos. I plan to test it on some real tasks – WordPress admin work, data migration, form filling – and write a follow up with what actually works and what doesn’t.

Feel free to give any thoughts on this.