😺Anthropic launched Fable 5.1: and now, the agents cost less

Anthropic has shipped Claude Fable 5.1, and the pitch is pretty clear: make agent work cheaper, and make the safety system stop tripping over ordinary tasks so often.

The biggest cost cut is on cache reads, which are used when an agent reuses context instead of reading it all again. Those now cost 75% less, down to $0.25 per million tokens. Anthropic says typical Fable workloads are about 25% cheaper overall, while highly agentic ones can save up to roughly 45%.

This matters most when the model is doing long, messy work. Anthropic says Fable 5.1 is better at long-running agent tasks, and its own agentic science test jumped to 52.6% from Fable 5’s 24.7%. The company also says biology safeguards interrupt benign requests 85% less often, while Claude Code sees about 60% fewer cyber-safety interventions.

The launch partners are leaning hard into the “leave it alone” story. They say the model completed a 38-hour machine-learning experiment unattended, solved a software crash that had gone unexplained for years, and finished 82% of Browserbase’s hardest agent tasks, up from 57% for Fable 5. Those are partner tests, not independent ones, so they’re useful but not the final word.

Anthropic also published a prompting guide that reads like a checklist for people who actually want these agents to work. Keep conversation history append-only. Re-test effort settings from scratch. Ask for progress updates. Batch tool calls. Tell the model to finish the whole assignment and not go off improving nearby bugs for fun. The message underneath all of it is simple: Fable 5.1 is less about raw cleverness and more about being worth trusting for longer stretches of real work.