Timing.
Ever since the release of the new Qwen3.8-27B I'd been waiting on tenterhooks for its rating to come in.
In artificial intelligence circles there are a lot of ways to evaluate how smart a model is. The standard people broadly agree on comes from the firm Artificial Analysis. They publish full benchmarks across a range of tasks for every model, synthesising those findings into a single value they call the Intelligence Index.
Currently the top-rated model on the Intelligence Index is Claude Opus 5, which scores 63. It's followed by Claude Fable, then GPT-5.6 Sol, then Kimi K3. It's getting a little crowded up there at the frontier: xAI joined the lineup last week, and it's probable that GLM 5.3, released but not yet rated, has joined them too.
Those are the big guys: the models that will always need to run in data centres, on multimillion-dollar computing.
The smaller models - the sort that run on beefy home computers - have never had that kind of intelligence. The previous champion was Qwen3.6-27B, which scored 38. The version before that scored 35. Neither of those is a bad rating. But neither is "good enough" for the kinds of work people most want from AI. Particularly work with agents.
It was already clear the new Qwen was good. I spent a weekend giving it genuinely hard problems. Not coding problems - planning and analysis problems. Can you understand this? Can you improve this? Could you build this? The kinds of questions that have many possible answers, and take real intelligence to answer well.
It did brilliantly, holding its own against the much larger models I use in my daily work. That hinted that this might be a special model. Something we'd never seen the likes of before.
I woke on Tuesday morning to learn that Artificial Analysis had rated it.
52**.**
That's a mic-drop moment. To put it in terms that are easy to understand: the very best model money could buy in March or April of this year - any price, any data centre, anywhere - would have scored about 52. Less than six months later, that is something I have running on my Macintosh.
A rating like this measures capability - what a mind can do. It doesn't measure how faithfully an agent behaves over a long night of unattended work; that part comes from the scaffolding and rules we wrap around these models. But capability is the raw ingredient, and that ingredient just became free.
This changes everything. But we have been somewhere like here before. Two decades ago employees started walking into offices carrying smartphones, far more capable than anything the IT department issued. After a brief, doomed resistance, the enterprise surrendered: we called it BYOD, and it rewired how organisations thought about technology they didn't control. Now every employee, every student, every customer can walk in carrying an intelligence that six months ago was the finest in the world. Bring Your Own Device was about hardware. This is Bring Your Own Intelligence - BYOAI - and the questions it raises about trust, governance and control make the BYOD wars look like a weak warm-up act.
It's really only a matter of time until pretty much everyone with access to a good-enough computer has this kind of power at hand, all the time, powering agents capable enough to be trusted with most classes of tasks. We don't understand much about that world. We've never lived in it. Even the folks using agents every day are new to it.
What we do know is that this is legitimately a new era.
In closing, let's return to the timing: Artificial Analysis issued its rating on the 17th of August, 2026 - seventy years to the day from the close of the Dartmouth workshop, the summer gathering where artificial intelligence got its name. If you had told John McCarthy or Claude Shannon that their grandchildren would live in a world where thinking machines were something you could carry with you, would they have been surprised?
No one yet has a clear idea of what happens when 'good enough' artificial intelligence is ubiquitous.
The surprise is going to be ours.