Frontier vs. open-weight model split

Why companies are becoming a series of loops | Anish Acharya (a16z)

Topic 5 of 19

Published Sep 6, 2026 · Clip 21:42–26:22 (4:40)

Anish explains the coming split in model choice through Pareto efficiency: frontier models like 'Mythos' and 'Fable 5' are irrationally priced — 100x more per extra IQ point than Opus 4.8 — but jobs with unbounded upside like drug discovery justify paying for peak intelligence. Bounded-upside functions such as legal, finance, and closing the books suit cheaper, RL-tuned open-weight 'mid-IQ' models, though a customer-support bug can be the first breadcrumb to a company-changing insight, making the case for frontier intelligence on some baskets of problems. He weighs the mid-IQ counterargument — we've crossed the intelligence threshold for nearly every economically useful problem — and notes we prematurely discard models that were the frontier just months ago.