Claude Fable 5 launches with Opus 4.8 fallback and 2.1.170 support

• Anthropic opened Claude Fable 5 across Claude Code, Desktop, Cowork, and the API, and ClaudeDevs' access note said Claude Code users may need version 2.1.170 before /model claude-fable-5 appears.
• The core product twist is that some cyber, bio, chemistry, and distillation prompts get rerouted to Opus 4.8, according to ClaudeDevs' fallback thread and Anthropic's launch post.
• Anthropic is pitching Fable as a longer-horizon coding model with always-on thinking, simpler prompting, and higher default effort, as ClaudeDevs' getting-started thread and ClaudeDevs' team workflow post put it.
• First-day creators used it for browser games, city simulators, and codebase audits, with bilawalsidhu's city simulator, GlennHasABeard's platformer, and marckohlbrugge's 37signals skills post showing the spread.
• Demand hit fast enough that people were burning through caps within minutes, and ClaudeDevs' rate-limit reset shows Anthropic reset both 5-hour and weekly quotas the same day.

You can read Anthropic's official launch post, the API team's model intro, and the new refusals and fallback docs. The weird bit is the billing path, where Anthropic also published a fallback billing cookbook. Then there is the early-user split: Every's week-long vibe check called it a warp drive for power users, while first-day X posts mostly alternated between one-shot demos and quota panic.

Anthropic's public package is straightforward: Fable 5 is the generally available Mythos-class model, while Mythos 5 stays limited to Project Glasswing and other restricted channels, per the launch post and the API model intro.

In product terms, day-one access broke down like this:

• Claude Code: /model claude-fable-5, per ClaudeDevs' getting-started thread
• Claude Desktop: update the app, per ClaudeDevs' access note
• Cowork: live on launch, per bcherny's availability post
• API and Managed Agents: switch the model ID to claude-fable-5, per ClaudeDevs' getting-started thread
• Pricing: $10 per million input tokens and $50 per million output tokens, per Anthropic's launch post
• Specs: 1M token context window and up to 128k output tokens, per the API model intro

The defining launch mechanic is the fallback path. According to ClaudeDevs' fallback thread, Fable 5 ships with new safety classifiers, and when one of them declines a request the system reruns that turn on Opus 4.8 instead of ending the session.

Anthropic's docs add three details that matter for anyone wiring this into tools:

• A refusal comes back as a normal API response with stop_reason: "refusal", not an HTTP error, per the refusals and fallback docs
• Server-side fallback can retry within the same round trip, per ClaudeDevs' fallback thread
• Anthropic also shipped client-side refusal middleware for Python, TypeScript, Go, Java, and C#, per ClaudeDevs' fallback thread
• The launch post says these classifiers fire in less than 5% of sessions on average, per the launch post
• Claude's cookbook says billing was changed so fallback usually does not make customers pay full token costs twice, per the fallback billing cookbook

Anthropic is selling a workflow change as much as a model. In ClaudeDevs' getting-started thread, the company says thinking is always on, responses can take longer, and effort controls how much thinking the model does, with “high” positioned as the default.

The more interesting claim is that older Claude habits may now be counterproductive. According to ClaudeDevs' getting-started thread, prior prompts and skills can feel too prescriptive for Fable, while ClaudeDevs' team workflow post says the internal shift was from checking whether Claude did the work right to checking whether it chose the right work.

That same thread maps the new default workflow into a few concrete moves:

• Start with harder tasks than you would have handed older models, per ClaudeDevs' getting-started thread
• Let Fable interview you for the spec, then run in auto mode overnight, per ClaudeDevs' getting-started thread
• Keep explicit success criteria through /goal or Outcomes, per ClaudeDevs' reroute note
• Use multi-agent orchestration to delegate subwork to smaller models, per ClaudeDevs' reroute note

Boris Cherny, Claude Code lead at Anthropic, pushed the same line in bcherny's workflow thread, where he described Fable as more methodical about measuring, logging, and verifying fixes before declaring victory.

The first wave of evidence looked less like benchmark chest-thumping and more like browser-native toy factories. bilawalsidhu's city simulator showed a city block simulator with multi-agent traffic, live detection boxes, and a day-night cycle, while bilawalsidhu's follow-up reply said the setup was just a /goal plus a “chonky well defined prompt.”

Other early examples clustered around lightweight games and visual prototypes:

• A 2D platformer made from an image reference and a simple prompt, per GlennHasABeard's platformer
• A simple fighting game using reference images for characters, per GlennHasABeard's fighting game
• An AI-startup Monopoly clone with rules, money, turns, and multiplayer share codes, per venturetwins' Monopoly post
• An F-Zero-style racer that needed a few extra prompts but still landed as a workable demo, per petergyang's F-Zero prompt post

The bigger pattern is that Fable's first-day wins skewed toward long-form code generation with visible payoff. minchoi's examples roundup turned that into a catalog of ten demos, from a Replit-style app builder to a browser Minecraft clone.

Not every useful demo was a one-shot game. marckohlbrugge's 37signals skills post used Fable 5 to read 37signals' open source Rails code and turn the patterns into reusable agent skills, which were then published as a GitHub repo.

That matters because it points at a different Fable workflow: using the model to compress a codebase into a reusable house style. The repo package is not just generated code, it is generated guidance for later code.

That codebase-scale angle also shows up in community positioning around Cursor. thekitze's Cursor test post kicked Fable loose on the @supermac_app codebase overnight, and thekitze's subagent post immediately jumped to orchestration, suggesting Codex could steer Fable subagents through the Cursor CLI.

The launch also had the familiar frontier-model messiness. ClaudeDevs' access note told users to update Claude Code to 2.1.170, while replies from Boris Cherny in bcherny's /model fable reply and bcherny's restart reply show people dealing with config flush delays and temporarily using /model fable as a shortcut.

The harder problem was burn rate. Dan Shipper's Every writeup said Fable could use 500,000 to 1 million tokens on tasks and priced it as roughly twice Opus. On X, LLMJunky's usage-limit post said Mythos one-shotted his usage limits, thekitze's limit joke said his limits were gone before he tried it, and ericzakariasson's Cursor post called it “incredible but expensive.”

Anthropic responded fast enough to make the scramble part of the story. ClaudeDevs' rate-limit reset says the company reset both 5-hour and weekly caps for all users the same day, and bcherny's reset reply shows support replies doing ad hoc resets for individual users too.

The buried caveat is data retention. The API model intro says both Fable 5 and Mythos 5 are Covered Models with mandatory 30-day retention, and they are not available under zero data retention.

That surfaced quickly in community notes. alliekmiller's retention note called out that zero-data-retention workspaces in Claude Console cannot use Mythos-class models without the 30-day policy, and Everlier's retention post highlighted the same requirement for Fable 5, Mythos 5, and future Mythos-class releases.

For a launch framed around capability, that is the part most likely to quietly decide where Fable can and cannot land inside bigger companies.