The Efficient Frontier

The Claude 5 generation has arrived, and most of the discussion has been about capability. Claude Fable 5 is currently the most capable generally-available model; part of a new tier, informally "Mythos-class," that sits above the Opus line. It joins a tight frontier cluster alongside Opus 4.x, GPT-5.5, and Gemini 3.1. The capability story is real. It is also, for most enterprises, the less important one.

The more consequential shift this generation is on the cost axis. Frontier-grade inference is getting materially cheaper, and the price of a given level of capability has fallen sharply over the past eighteen months. Falling token costs do more than trim the bill; they change what is economically viable. Workloads that were uneconomical a year ago, including always-on agents, long-running reasoning loops, and putting an entire corpus in context instead of retrieving from it, are now defensible line items.

This reframes the question every platform team is asking. It is no longer "which model is best." It is "which point on the capability-versus-cost curve fits this workload." That curve, the efficient frontier, is the organizing idea of this paper.

What follows: what the Claude 5 generation changes, why cheaper inference matters more than another benchmark point, how to treat capability tiers as an architecture decision rather than a procurement one, and a discipline for adopting a new model generation without quietly destabilizing the systems you already run in production. The two most expensive mistakes we see in the field, over-paying for intelligence on trivial work and upgrading models without re-running evals, are both avoidable with the framework here.