GPT-6 Astra can finally see the interface it is building

OpenAI shipped GPT-6 Astra on 3 September, as a limited preview for trusted partners, and began widening it to paying ChatGPT users over the following days. It is their most capable model so far: a 1,050,000-token context window, up to 128,000 completion tokens, and a knowledge cutoff of 30 April 2026.

Most of the coverage has gone to the safety story — it is the first model OpenAI has rated at the Critical level for cybersecurity capability under their Preparedness Framework, and they held it back roughly four weeks over exactly that. That is the interesting story if you work in security. It is not the interesting story if you build interfaces.

The number nobody put in the headline

On ScreenSpot-Pro — a benchmark that asks a model to point at a specific element on a screen, with no tooling to help it — Astra scores 92.7%. GPT-5.6 Sol scored 76.9%.

Fifteen points is a lot, but the size of the jump is not really the point. The point is what sits on either side of the line. At 77% a model that is asked to click the submit button gets it wrong roughly one time in four — which means you cannot let it act unsupervised, and the whole loop stays manual. At 93% it is wrong closer to one time in fourteen. That is the difference between a demo and something you would actually leave running against a staging build.

OpenAI's own framing is the giveaway: Astra can build a website and then run front-end QA checks on it to confirm the features work. Not generate markup and hand it over — open it, look at it, and verify. Every solo builder I know has the same hole in their process, and it is exactly that one.

On the speed claims

This is where I would slow down. The figures being quoted are real, but they are not all the same kind of number.

• In a latency simulation on OSWorld 2.0, Astra finished tasks in about 40 minutes against roughly 75 for Sol — a reduction of around 47%.
• Output speed tops out around 63 tokens per second on the fastest configuration.
• Time to first token bottoms out around 2.10 seconds on the low-reasoning setting.

The first is a benchmark result. The second and third are measured throughput. And reviewers with early access reported that the model simply felt unusually quick — while noting, to their credit, that the early tester pool was small and the servers were barely loaded.

I would treat the speed story as unproven until then. The capability story does not depend on it.

What it costs

$10 per million input tokens, $50 per million output. Cached input drops to $1. Batch and flex processing run at half rate. There is a Fast mode at roughly double the price for up to 2.5× the speed, and web search is billed separately at $10 per thousand calls.

For a solo operation that is not nothing. Output at $50 per million is the line that will decide whether you use this in a loop or reach for it deliberately. My guess is most people will end up doing what I do — a cheaper model for the hundred small edits, and the expensive one for the thing that would otherwise cost an afternoon.

Getting hold of it

In the API it is gpt-6-astra, and it is also going out through Microsoft Azure and Amazon Bedrock. Pro, Business and Enterprise plans additionally get Astra Pro. The rollout has been staggered rather than clean — OpenAI opened it to most paying users a day after unveiling it, and developer access lagged behind the consumer apps, which drew a public apology for the mess.

So: has it taken over the internet?

Not yet, and it would be strange if it had — it has been generally available for a matter of days. What has happened is narrower and, I think, more interesting. A model got substantially better at the one thing that has been blocking automated front-end work: looking at a rendered interface and knowing what it is looking at.

I will write again about this once I have run it against something real — probably a QA pass over one of my own dashboards, which has more edge cases than any benchmark. Until then the honest summary is that the capability numbers look like a genuine step, and the speed numbers need a month of real load before they mean anything.