This week OpenAI launched GPT-6 Astra and called it the most intelligent, most aligned model ever built. Its president talked about a new generation of intelligence. Asked point blank whether this counts as AGI, he said that, personally, he thinks so.
The numbers behind the launch are big. One reasoning benchmark jumped from 8% to nearly 100% in a single model generation. An advanced math test scored 98%. An offensive-security test scored 100%.
And yet, on a broader index that measures general intelligence — not one benchmark picked for the headline — GPT-6 Astra ranked behind several competing models. Usage pricing rose two and a half times versus the previous model. Broad access still isn’t available: for now it’s limited to a small group of organizations.
None of those three facts made the headline. All three matter more than the headline.
This happens every week
This isn’t the first time, and it won’t be the last. The same week Astra launched, Google shipped a weather model that halves rain-forecast error, Nvidia acquired an AI platform for over $12 billion, and a competing model closed a math problem in a day that had stood open for years. Every week brings an announcement that “changes everything.”
For a business that bills, collects, and delivers every single day, keeping pace with every launch is a full-time job nobody has time for. The good news: you don’t need to.
The model isn’t the decision that matters
No model, however many benchmarks it wins, tells you which information is a mess, which decision repeats every week without needing to, or where your team loses hours waiting for an answer that already existed somewhere. You have to map that yourself, with business judgment, before picking which AI solves it.
A well-designed decision system doesn’t depend on which model is trending this week. It’s built so the model underneath is swappable: one today, whichever wins the next benchmark tomorrow. What doesn’t change is the judgment behind what you chose to automate.
What this means for your business
Before evaluating which AI model is “the best,” answer a smaller question first: which decision do you make every week without a clear criterion? That’s the one to solve first, whatever model you use. Everything else is launch noise.