The Efficient Frontier

Anthropic announced Claude Fable 5 and the restricted Claude Mythos 5 on 9 June 2026. The official release and access updates also record a suspension on 12 June and restoration on 1 July. Access was therefore suspended when this paper was first published on 20 June. The Fable model page subsequently introduced Fable 5.1 on 1 September. These are dated release facts, not a timeless ranking of the best model.

For a platform team, the useful question is how a model performs at an acceptable cost on its own workload. Token prices, retries, cache behavior, latency, tool calls and human review all affect that answer. A lower advertised token price does not establish a lower cost per accepted result or make an always-on agent economical.

This paper uses the efficient frontier as a planning model: compare candidates on task-specific quality and total cost, then choose the combinations that satisfy the workload’s constraints. A benchmark can inform a candidate list; it cannot replace that comparison.

The following sections cover workload-specific model selection, optional capability tiers, regression evaluation and staged adoption. They provide planning practices for reviewing over-provisioning and upgrades without evaluation.