Pleias co-founder Pierre-Carl Langlais argues frontier models like Fable face severe economic hurdles over inference margins and alignment

Gpt sol where the sol short for "solution" (to the Dario problem)

@Dorialexander Gpt sol

@Dorialexander If gpt-5.6 sol ultra is a fable level model then openai might have an edge there.

@Dorialexander the trend suggests we use models such as fable 5 for ideation/prototyping and then once we are ready to deploy agents in applications, we switch to the strongest open-source models such as glm 5.2

using fable 5 in production would be unsustainable

@yacineMTB I heard Shit Outta Luck and no other acronym makes sense anymore

@Dorialexander capability and unit economics decouple at every launch. a frontier model ships as a loss leader until a quant pass and speculative decode drag serving cost back under the price they already committed to.

@yacineMTB Is it ... A final solution?

@Dorialexander fable is the right model to benchmark against, sonnet is the right model to ship.

@yacineMTB awkward spot between capability and viability is where every frontier model lives for a few months before the optimization catches up

Wonder what proportion of fable's costs are running every shred of i/o through layers classifier models. Once they're tuned and distilled, once the next 'safer', 'smarter' (on some tasks) checkpoint is decanted, and once the feed of sauron's pr-value boosting eye gets impresssed with other newer shiny things, they'll lean into quietly making inference optimisations.

But they're also def testing the water to gather data for revenue projection impacts from funneling customers into paying retail API rate fees for accessing a model that (seems to) understand the assignment, completes it, and doesn't lie like it's running for office while being a preachy asshole.

@Dorialexander Open weights change the math. GLM hits ~95% of frontier at ~80% less cost with no alignment tax in serving.