Breunig says slower model gains make harness work pay off

Drew Breunig argued that Fable's release ends a free ride AI engineers have had for years, in a post published August 23. The stakes are concrete for anyone routing work across models. GLM 5.2 launched the same month at about a ninth of Fable's price and a fifth of Opus 5's, per Breunig.

Sutter's 2005 essay, 'The Free Lunch Is Over,' supplies the comparison. CPU clock speeds doubled every 18 months for decades, so optimizing code was a bad bet. Faster hardware would fix a slow program within a year or two.

When clock scaling stalled, optimization stopped being optional.

The model-release cycle did the same job for years, he says. Tuning a harness or a context strategy didn't pay off, because a cheaper, better model would show up in a few months. It would cover for whatever you'd built.

Fable changed that math. It's strong and priced high enough that teams now have to decide what work goes where. Waiting for the next release to close the gap doesn't work anymore.

His workflow shows how it works. Fable shapes the design, then a written brief goes to GLM 5.2 to execute it.

Its access controls and data-retention rules will lock this multi-model split in, Breunig expects, since companies will start caring where their traces go.

Ramp's numbers already show developers running a portfolio of models instead of one default. Breunig's post argues the reason is that price-per-dollar model gains slowed, so the payoff shifted to everything built around the model.

Breunig doesn't know if the split holds past the next release cycle, and says as much in the post.