Anthropic's Answer to Fable 5's Price: Let It Manage Cheaper Models
Two official patterns, Advisor and Orchestrator, deliver up to 96 percent of Fable 5's performance at roughly half the cost by pairing it with Sonnet 5.
Claude Fable 5 is the strongest model Anthropic sells, and one of the most expensive, at $10 per million input tokens and $50 per million output tokens. Anthropic’s new advice to developers: don’t let it do all the work. In official cookbook examples and documentation, the company now recommends using Fable 5 as a kind of senior colleague that guides cheaper models instead of typing every line itself.
Two patterns do the trick. In the “Advisor” pattern, the smaller Sonnet 5 does the actual work and only calls Fable 5 when it gets stuck, roughly once per task, according to Anthropic. On SWE-bench Pro, a benchmark of real software-engineering problems, this combination reaches about 92 percent of Fable 5’s solo performance at 63 percent of the cost. The second pattern flips the hierarchy: Fable 5 acts as an “Orchestrator” that makes the plan and delegates the pieces to multiple Sonnet 5 workers. On BrowseComp, a web-research benchmark, that setup delivers 96 percent of the performance at 46 percent of the cost. Both run through Claude Managed Agents, with each sub-agent keeping its own memory cache to avoid paying for duplicated context.
Here’s what’s behind it: price pressure, from every direction. Chinese open models like GLM 5.2 undercut Western prices dramatically, xAI’s brand-new Grok 4.5 costs a fifth of Fable 5, and OpenAI’s GPT-5.6 launches cheaper and reportedly more token-efficient too. Anthropic’s counter isn’t a price cut, it’s a usage manual for spending less. The honest reading: “plan big, execute small” is genuinely good engineering advice that teams have used informally for a while. Now it’s official, documented, and benchmarked.
What this means for you: If you use Claude through the chat app, nothing changes, this is for people building with the API. If that’s you, the pattern is worth stealing even outside Claude: use an expensive model for judgment and planning, cheap models for execution. The same logic applies to any provider’s lineup. And if you’re just watching from the sidelines, there’s a nice takeaway about where AI is heading: the future probably isn’t one giant model doing everything, but teams of models, a smart one managing, cheaper ones doing, which is, for better or worse, a very familiar org chart.
Sources
Source: https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool
Meta Is Testing AI Glasses That Record Your Whole Day, Without a Recording Light
A prototype feature called Super Sensing captures continuous audio and photos every few seconds so an AI can recall anything you saw or heard, the Financial Times reports.
