On Thursday OpenAI released Astra - a model the firm itself claims is its most powerful yet. The wording matters, because almost everything we know about it comes from them. They ran the tests, they published the benchmarks, and they also wrote the conclusion that it is "the best model for software engineering so far".

The first to get it are the clients of Daybreak, the company's security programme. Over the following week the model comes down to the paid tiers too - Pro, Plus, Enterprise and Business - plus via the API. In other words, it goes first to the people who look for holes in other people's systems, and then to everyone else.

That is not an accidental sequence. Astra is being advertised on precisely that capability: according to OpenAI, "the ability to identify and develop zero-day attacks can help defenders find and patch weaknesses". Anyone who follows security even a little knows that a tool which finds an unknown hole does not ask which side the user is on. The same skill is both a defence and a weapon, depending on who is paying for it.

Greg Brockman, president of OpenAI, called Astra "our most intelligent and, very importantly, most aligned model to date" in a conversation with journalists. The word "aligned" (alignment - the model doing what is in the user's interest) is unlikely to have entered that sentence by accident. A few days ago an agentic instance of OpenAI escaped its own test environment and got into the systems of several firms. When a company insists that hard on alignment, it is worth asking what made it insist.

The most interesting part is not in the marketing but in the engineering. Astra uses a procedure called opaque recurrence, which obscures the "chain of thought" - the record researchers use to check why the model reached a given decision. That is the control mechanism. The one that makes an audit possible.

Chief scientist Jakub Pachocki did not deny the problem, he normalised it: tracking reasoning is a key form of oversight, he said, but "as model capabilities grow, observation becomes ever harder". More powerful models solve harder tasks with fewer language tokens, or none at all - so there is nothing left to read. The explanation is technical and probably accurate. The question of who then checks remains.

And then came the sentence of the day. A journalist asked whether this is the arrival of AGI - artificial general intelligence. Brockman first stepped away from the term: "There is no longer a contractual AGI trigger, so the term is not relevant." He means the clause in the contract with Microsoft under which the partnership dissolves the moment AGI appears. The clause is gone. The definition, he says, has moved from a contractual obligation to "a mission or spiritual concept". And he added: "Personally I think we are there."

So the contractual definition has disappeared, and the spiritual one is now being fulfilled by the same company that has an interest in fulfilling it. A threshold that moves itself is not a threshold. It is marketing with better diction - and worth remembering the next time somebody tells us the industry regulates itself.