Many people are finding that the advantages of Fable often aren’t worth the cost. (See data below for a sample analysis.)
This doesn’t mean that Fable doesn’t have its uses – clearly it does.
But it poses a *serious* problem for massive data center investments.
It suggests that newer, immense models will be niche that customers usually won’t pay for. Even bigger models might not be worth building at all, economically, because if they are even more expensive even fewer user will use them.
For a lot of purposes open models will suffice. (For some purposes in which the cost of error is high, no models may suffice.) In some cases people may not even need cloud, and prefer on-premie, especially as models become more optimized for local use.
If the above scenario is true, you have to ask whether we need quite so many data centers.
If we don’t, we end up with an oversupply.
The fallout across the economy is likely be immense.
@GaryMarcus And it’s even worse when you’re measuring cost to run vs price. 130% more than opus 4.8.
4.5x GPT 5.5 high for ~10% higher benchmarks.
Benches aren’t perfect but the share of problems where even a 30% better result is worth 4.5x more is not large…
