The headline number in the fermisense writeup is easy to misread: a GRPO fine-tune of a 9B open-source model hitting $0.50 per 1,000 catalog listings, roughly 40× cheaper than the least expensive frontier setup and ~340× cheaper than the priciest. The interesting part isn’t that a small model won — it’s the conditions under which it won. Same workflow, same tools, same images, same scorer. Hold everything constant and the only variable left is who owns the weights.

That’s the shift worth naming. Frontier models are priced for generality you mostly don’t use. A catalog-integrity task has a narrow decision surface — is this listing correct, complete, mislabeled? — and reinforcement fine-tuning lets a 9B model memorize that surface instead of reasoning its way there on every call. In the production AI work I’ve done, the cost that quietly wrecks a budget is rarely the hard case; it’s the per-item tax on the millions of easy ones. A task-trained model collapses that tax.

The catch nobody puts on the cost chart is the eval harness. GRPO is only as good as the scorer that shapes the reward, and building a scorer that actually correlates with catalog quality is most of the real work — the fine-tune is the cheap part. The HN thread has the predictable “but does it generalize?” pushback, which misses the point: you’re not buying generalization, you’re buying a specialist and keeping the weights.

If owning task-specific models at $0.50/1k becomes routine, the frontier labs’ moat stops being raw capability and starts being the tasks nobody has bothered to specialize yet. How many of your LLM line items are paying frontier prices for a job a fine-tuned 9B would do better?