Anthropic's Claude Fable 5 launches with 9-hour sustained runs and top benchmark scores

Claude Fable 5 Launches With Extended Reasoning and Lower Per-Token Costs

Anthropic has released Claude Fable 5, a model designed around sustained inference rather than raw capability gains. The headline differentiator is endurance: users report running the model for nine hours straight without performance degradation, a capability aimed at long-horizon reasoning tasks.

Early reaction splits along a familiar line. The Information's Dan Shipper characterizes it as safer and more reliable than previous versions, though frames it as a trade-off—less maximum capability in exchange for consistency and lower cost. Pricing reflects this positioning: $10 per million input tokens and $50 per million output tokens, making it cheaper to run at volume than competing models.

The inference compute thesis

What's notable in the rollout isn't the model itself but the worldview it signals. Noam Brown published analysis on large-scale test-time compute suggesting that performance plateaus, when they appear at all, sit far beyond practical budgets. The implication is that inference—the act of running a model on a task—is becoming the real locus of compute spend and competitive advantage, not training.

André Karpathy observed this in auto-research experiments where performance continued improving even after hundreds of runs. The extreme case Brown raises is conceptual but real: a deployed model might still be running a task when the next model version launches. This is a different scaling paradigm than the industry has operated under—one where you don't ship a model and move on, but rather ship a model and let it think for hours or days while you prepare the next iteration.

Whether Fable 5's nine-hour sustained runs are engineered specifically to demonstrate this capability, or whether they're a side effect of the safety-focused design, the message is the same: reasoning time, not parameter count, is where the next edge lies.