Anthropic updates Fable prompting with /model, high effort, and /goal loops

• Anthropic's launch guidance for Claude Fable 5 starts with a literal model switch, because ClaudeDevs' setup thread says Claude Code users should run /model claude-fable-5, while the API now expects the claude-fable-5 model ID in both Messages and Claude Managed Agents.
• Fable's default operating mode is slower and more deliberative, because ClaudeDevs' setup thread says thinking is always on and recommends high effort as the default, with even low and medium often beating prior models at xhigh in Anthropic's evals.
• Anthropic is telling teams to delete scaffolding, not add more, because ClaudeDevs' setup thread says older prompts and skills can be too prescriptive for Fable, and ClaudeDevs on team workflow says the internal shift is from checking whether Claude did the work right to checking whether it is doing the right work.
• Self-verification moved from nice-to-have to core harness design, because Boris Cherny on self-verification calls it a key ingredient for longer autonomous runs, and Matt Pocock's early Fable workflow shows /goal and autoCompactWindow already showing up in day-one usage.
• Fable's safety layer is part of the product contract, because ClaudeDevs' setup thread says some cyber and biosecurity requests auto-reroute to Opus 4.8 at Opus pricing, while Anthropic's fallback billing guide adds server-side fallback and token-credit handling for API users.

You can read Anthropic's launch post, jump from the general prompting best practices page to the Fable-specific guidance it references, and check the separate fallback billing guide for the part the announcement barely sketches. The most useful change is not a new command, it is the prompt philosophy: the Claude Code team's workflow note and Boris Cherny's self-verification post both describe longer runs where the harness verifies goals and progress instead of micromanaging every step.

Anthropic's first instruction is mechanical. Claude Code gets Fable through /model claude-fable-5, and API users switch the model ID to claude-fable-5 in both Messages and Claude Managed Agents, according to ClaudeDevs' setup thread.

Access was patchy enough on launch day that Anthropic posted a separate fix note. ClaudeDevs' access reply says Claude Code users should upgrade to CLI version 2.1.170, while Claude Desktop users need the latest app version.

Anthropic's Fable guidance breaks with the usual "more scaffolding is safer" instinct. ClaudeDevs' setup thread says thinking is always on, responses can take longer, and high should be the default effort, with xhigh reserved for the hardest tasks.

The more surprising line is the prompt migration advice. According to ClaudeDevs' setup thread, existing prompts and skills built for older Claude models are often too prescriptive for Fable, which is why Anthropic's prompting best practices page now points developers to a separate Fable-specific page covering effort levels, long-run progress claims, memory systems, and instruction changes.

Inside the Claude Code team, ClaudeDevs on team workflow frames that shift as moving from verifying whether Claude did the work right to verifying whether it is doing the right work.

The biggest recurring theme across Anthropic's own posts is verification loops. ClaudeDevs' setup thread says feedback loops still hinge on giving Fable explicit success criteria, either through /goal in Claude Code or Outcomes in Claude Managed Agents.

Boris Cherny, Claude Code lead at Anthropic, made the same point more directly in his self-verification post, writing that powerful models running for long periods need self-verification loops to stay closer to the intended result.

Early usage looks a lot like that advice:

• Matt Pocock's early Fable workflow used /to-prd to create a spec, then /goal to implement against it.
• The same Matt Pocock thread set "autoCompactWindow": 180000 to keep the session in the "smart zone."
• Lance Martin's loop-design post separately pointed people to self-correction loops and memory as the two knobs worth tuning first.

Fable's safety behavior is not just a footnote in the launch thread. ClaudeDevs' setup thread says cyber and biosecurity requests may auto-reroute to Opus 4.8, shown in the UI and billed at Opus prices, while Anthropic's launch post says those safeguards trigger in under 5 percent of sessions on average.

Anthropic published a separate classifier fallback and billing guide the same day with more operational detail:

• API users can configure built-in server-side fallback from Fable 5 to Opus 4.8.
• Clients not using server-side fallback need to handle a fallback_credit_token in stop_details for billing.
• Anthropic says most fallback cases should avoid charging users for Fable tokens.

The other operational clue came from Boris Cherny's /usage note, which says Claude Code's /usage view now breaks down token spend by skills, MCPs, and plugins.