Roupen Odabashian, Oncologist at Abbotsford Regional Hospital and Cancer Centre, Founder of MeDucation AI, and Podcast Host at OncoDaily, shared on LinkedIn:
“Everyone still talks about AI progress in terms of parameter count, bigger model, more compute, better answers.
That framing is getting stale. Some of the sharpest recent gains aren’t coming from scale at all.
They’re coming from architecture: how a model checks its own work, how it interacts with tools and its environment, how it decides when to think harder versus when to just answer.
Anthropic’s new Claude Opus 5, makes that shift visible. It ships with a dial that lets you trade cost for capability in real time: crank it down for a quick draft, crank it up when you need it to actually think.
Anthropic says it approaches their top tier model’s intelligence at roughly half the price, with real gains in coding and multi-step agent work.
None of that came from a bigger model. It came from something they call self-verification, the model checking its own output before handing it back.
That’s the part worth sitting with. A model that checks its own work is a fundamentally different object than one that just produces output fast, even if the underlying parameter count didn’t move.
It changes what “close enough” means for anything with real stakes.
Most people are still using these tools like a vending machine: press button, get answer.
The effort dial is a small UI change that implies something bigger, that the right amount of verification is a variable you’re now responsible for setting, not something baked into raw model size.
What does it look like when professionals in high stakes fields (medicine, law, engineering) actually learn to use that dial well, instead of assuming more parameters is where the next gain will come from?”
