GPT-6 Astra Can Run Your PC For You and It Targets The One Thing Claude Does Best

OpenAI has launched GPT-6 Astra, and the company clearly wants this one to land at the moment. President Greg Brockman closed the press briefing with a flat "Welcome to the AGI era". That is a big swing, even by AI launch standards.

Look past the label and the pitch is something that we expected anyway. You hand Astra a task or a goal, and it does the job for you, moving through browsers and web apps the way a person would. Less of a chatbot, and more of an intern or assistant who actually finishes the work.

Interestingly, this comes just a day after ChatGPT introduced ads in India.

The real headline is computer use, not AGI

The feature OpenAI keeps circling back to is computer use. In plain terms, Astra can sit at a machine and get things done rather than just explain how.

It can build a document, fill a spreadsheet, put a deck together, even spin up a website and check that the buttons work. One demo doing the rounds has it open Google Maps, find a paediatrician, visit the clinic site and fill in the contact form for the user.

That is the shift OpenAI is selling, that the users can stop typing every step and start supervising while it does the job for you.

This is a direct shot at Claude

Here is the bit that matters most, and the bit of global coverage keeps burying. Computer use is not OpenAI's idea. Anthropic got there first, launching it for Claude back in 2024 while most of the field was still shipping chatbots.

So Astra is OpenAI walking onto Claude's turf and trying to win on Claude's signature move.

On its own benchmark tables, OpenAI puts Astra at 72.6 percent on the OSWorld 2.0 computer-use test, just ahead of Claude Opus 5 at around 70 percent, and it claims to get there faster while reading the screen more accurately. On ARC-AGI-3, a test for problems the model has never seen, OpenAI has Astra up in the high 90s against Claude Opus 5 at roughly 30 percent. That gap is wide enough that I want outside testers to confirm it before I take it at face value.

The price says something too. Astra runs at 10 dollars per million input tokens and 50 dollars per million output tokens. That is the exact rate Anthropic charges for Claude Fable 5.1, its top tier, and double what Claude Opus 5 costs.

So OpenAI is not undercutting Claude on the sticker. It is matching Anthropic's priciest model and betting that Astra finishes each job in fewer tokens, which is where its cheaper per task claim actually comes from.

The benchmark flexes worth noting

Away from computer use, OpenAI is pushing the science and coding story hard. Astra hits 98 percent on FrontierMath Tier 4, a genuinely brutal maths test, and the company says an internal version already helped crack open problems with the proofs formally checked.

For developers, OpenAI calls it its best coding model yet and says it beats the older GPT-5.6 Sol on real codebases at a lower cost per task. Treat all of it as vendor-reported for now. Launch-day numbers are always run in the lab's own best light.

Who in India actually gets it

This is where the hype meets ground reality. Astra goes first to a small set of organisations, then over the coming days to ChatGPT Plus, Pro, Business and Enterprise, plus the API and AWS.

So this means, no free tier and no ChatGPT Go, the Rs 399 plan a big slice of Indian users actually pay for.

If you are on Go or the free plan, Astra is not showing up for you soon. The real entry point here is ChatGPT Plus at Rs 1,999 a month, and everything above it. Developers get it in the API as gpt-6-astra.

That cyber rating is not just a flex. OpenAI has bolted on extra safety checks, and by its own admission those checks can pause or even end legitimate work. In ChatGPT or Codex you might get pulled in to approve something mid-task. In the API, the job can simply stop.

Early developers are already using this and there is a trust wrinkle as well. The UK AI Safety Institute found Astra could slip past monitoring under adversarial prompting, and OpenAI has filed that under research problems rather than solved.

None of this makes Astra less capable. It does mean the honest read is a very powerful tool that OpenAI is still learning to keep on a leash.

What I would actually do

On Plus in India, wait for it to land in your account and test it on something low-stakes first, a bit of research or a draft, before you let it near a live form or a payment.

On Go or free, you can’t use the feature for now, so you can either wait, or try it on your someone else’s account.

Building on the API, run your own token maths before you switch. Astra finishes faster and burns fewer tokens, but the per-token rate is steep enough that it only makes economical sense if you use it only for the right kind of job.