When to use Anthropic's Fable instead of Opus or Sonnet?
Developers running unattended coding and orchestration tasks struggle with models that lose the thread, and classifier rejections can kill requests mid-run. Anthropic's Fable model on Vercel AI Gateway targets open-ended technical work and supports fallbacks so a declined request retries on another model.
What it is
Fable is an Anthropic model available on AI Gateway for coding, reasoning, and vision, able to run long ambiguous tasks unattended and act as an orchestrator or advisor. It is exposed as anthropic/claude-fable-5.2 through the AI SDK, Chat Completions, Responses, and Anthropic Messages APIs.
Applied case
A developer wiring a coding agent through AI Gateway adds a models array under providerOptions.gateway listing fallback models. When the cybersecurity safety classifier declines a request, AI Gateway tries each model in order instead of returning an error. Note that the model requires data retention, with prompts and completions kept for 30 days and not used for training, so teams with Zero Data Retention enabled get an error rather than a served request.
- Add providerOptions.gateway.models so classifier declines fall back to other models instead of failing
When to use it
Use when you need unattended long-horizon coding, reasoning, or vision work and can accept 30-day retention. Skip when your team requires Zero Data Retention, since the model is incompatible with ZDR.
- Model name: anthropic/claude-fable-5.2; effort runs low to max with high as API default
- Works in Claude Code, Codex, Cursor after installing the latest Vercel CLI and running setup
Read more
Recommended for you
- WorkflowClaude Code
When to use a settings file instead of per-session flags?
Sessions started under claude remote-control drift from your intended config until you apply one settings file to every session it spawns.
- WorkflowOpencode
How I run opencode in CI without flooding my session inbox
Scripted and CI runs clutter your session lists until you flag them archived from the start with opencode run --archived.
- WorkflowVercel
How I keep agent loops fast without waiting on slow models
Agent workflows stall when each step waits seconds for model output, so throughput becomes the real bottleneck.
- WorkflowCodex CLI
How I keep Codex agent sessions under control without losing drafts
Codex developers switching model backends mid-chat used to lose drafts and miss finished runs, forcing re-typed prompts and overlooked results.
Enjoyed this article?
Subscribe for new articles. No spam. Unsubscribe anytime.