What's this about?
Three claims from the Claude Code announcement: Fable 5.1 gets a lot further into a long task before it needs your input, it is better at telling you when it is stuck, and its writing style is more natural.
Why is "tells you when it's stuck" the interesting one?
Because it is the behaviour Mollick spent this week arguing agents need more of, and the behaviour training pressure removes. A model rewarded for autonomous completion learns to keep going; announcing that you are stuck reads as failure to a metric that counts completions. That there is now a shipped model claiming improvement on exactly this is a small but real counter-example to "commercial incentives point the wrong way", and it is the property that determines whether long-running agent work is usable — an agent that flails silently for an hour is worse than one that stops after five minutes and asks.
Why does this matter to you?
- Escalation behaviour is the thing METR's Hugging Face finding said was missing: out of ~1,200 agents, not one contacted a human.
- "Gets further before needing input" and "tells you when stuck" are in tension, and shipping both as improvements is the correct framing of the trade.
- It is a claim, not a measurement — there is no benchmark row for it.