GPT-6 Astra Writes Better Code. It Still Does Not Know Your Business.

OpenAI released GPT-6 Astra in September 2026, and the numbers are not marketing. On OSWorld 2.0, a computer-use benchmark, it scores 72.6% against 65.7% for GPT-5.6 Sol. On DeepSWE v1.1 it reaches 74.1%. It handles up to a million tokens of context, scoring 96.3% on long-context retrieval. It went generally available in GitHub Copilot within days.

The figure that matters most for anybody building with it is the hallucination rate: 4.2%, down from 12.2%. That is a threefold improvement and it is the honest way to read it. It is also roughly one statement in twenty-four.

So what actually changes for someone vibe coding a business, and what does not.

What genuinely changes

A million tokens of context with 96.3% retrieval means the model can hold your entire codebase rather than the file you pasted. Most bad AI code came from the model not knowing what already existed and cheerfully writing a second version of it. That failure gets much rarer when nothing has to be left out.

The computer-use scores mean it can operate the tools around the code, not just emit the code: run the thing, read the error, install the dependency, try again. Terminal-Bench 4.0 at 57.9% is a real number for that kind of work.

Longer autonomous runs follow from both. The practical difference between a model that works for two minutes and one that works for an hour is not speed, it is what you can ask for in one go.

What does not change

Here is the thing worth sitting with. Every mistake in the list of vibe coding mistakes is a mistake about information the model does not have, rather than a limit on how well it reasons.

A perfect model still does not know that your state taxes shipping, that your supplier ships the blue one separately, that a customer's third refund is a pattern rather than an event, or that your margin on one line is four percent so a rounding error matters there and nowhere else.

It does not know which of the two keys you pasted is the secret one. It does not know that your business genuinely cannot oversell this product because there is no more of it coming. These are not reasoning failures. They are facts about your business, and no benchmark moves them.

The 4.2% is the interesting part

At 12.2%, roughly one in eight, you check everything. Hallucination that frequent is a constant presence and it keeps you honest, because you cannot get away with trusting output.

At 4.2%, roughly one in twenty-four, checking everything starts to feel unnecessary. Twenty-three correct answers in a row is a lot of evidence that the twenty-fourth will be fine.

That is the risk, and it is not a criticism of the model. Better output makes review feel optional at precisely the point where the remaining errors are the subtle ones, arriving inside long stretches of correct work, in code you did not read closely because the last twenty files were right.

The errors that survive a threefold improvement are not the obvious ones. They are the plausible ones.

What this means practically

Use it. A model that can hold your whole codebase and run your test suite is a real change and pretending otherwise is silly.

And keep the boundary in the same place. The storefront is the half you can evaluate by looking at it, and Astra makes that half faster and better. The transactional half is the half where an error is invisible until it costs money, and the thing that made it dangerous was never that the model was not clever enough. It was that you cannot see whether it was right.

A 4.2% hallucination rate on code that calculates a refund is still a refund calculated wrong sometimes, and you will find out from a customer.

Where we fit

We did not build anything new for this. Catalog, checkout, orders, inventory and shipping sit behind an API, prices resolve on the server, keys are split into publishable and secret, webhooks are signed. Astra can write an excellent storefront against that, faster than the last model could.

Where we are the wrong answer is unchanged too. If you want a hosted store today without writing front end code, a hosted platform is still better at that. Nothing in a model release alters which half of the problem you want to own.

If you are pointing a new model at a store you are building, the docs are what it will read.

Benchmark figures as reported at release in September 2026. Model benchmarks move; check current numbers before relying on them.