The Thing
Anthropic shipped Claude Fable 5.1 on September 1st, and buried in the release notes is the actual headline: it’s a model built to be left alone for multiple hours or days. 1M-token context, whole-codebase changes, unattended agents, multi-app workflows, and cache reads cut 75%, from $1.00/M tokens to $0.25/M (TechSpot). Anthropic says that pricing change alone cuts effective cost by up to 45% on heavily agentic workloads.
Alongside it: Mythos 5.1, same weights, gated access, aimed at cybersecurity and biology work Anthropic thinks is too dangerous to hand out freely.
I’m asking myself now, “How long can I walk away while Fable builds for me?”
The Mechanics
Terminal-Bench 4.0 jumped from 42.0% to 55.8%. Terminal-Bench-Science more than doubled, 24.7% to 52.6% (CellCog). But the number that should make every engineering lead pay attention is buried in the pricing mechanics, not the benchmark table: cache reads. In a long-running agent, the model isn’t generating fresh tokens most of the time, it’s re-reading a growing context window on every single tool call, every file check, every test rerun. That re-read cost compounds fast, and it was the real tax on running agents unattended. Cutting it 75% is why Anthropic can plausibly claim a 45% cost reduction on the exact workload pattern that makes agents expensive: task, plan, tool call, test, recovery, verification, repeat, for hours.
There’s a tell in the Mythos numbers too. Same underlying weights as Fable 5.1, but running with more permissive safety settings, Mythos scored 60.9% on Terminal-Bench 4.0, five points higher than Fable 5.1 (tech-insider.org). Which means a meaningful slice of “how good is the model at coding” is actually “how often does the model get stopped from doing the coding.”
The Point
This is the latest big shift from AI-as-suggestion to AI-as-delegate, and the economics finally support it. A junior engineer costs a salary regardless of whether they’re grinding through a migration or waiting on a code review; an agent burning $50/M output tokens for three days unattended needed the cache-cost fix to make that math work at all. What changes: teams start measuring agents on completed-PR-per-dollar instead of code-quality-per-prompt. What should worry you: the safety-classifier tax is now a visible, quantifiable line item, which means someone, somewhere, is going to start optimizing for the wrong side of that tradeoff. The gap between what a model can do and what it’s allowed to do just became a a feasible 8 hour work day business decision.