# FableGuide

> A field guide to Claude Fable 5, Anthropic's most powerful model (released 2026-06-09). Model ID, pricing, availability, first API call, breaking changes, tuning guidance, and the Mythos 5 research story. Part of the WholeTech network. Built Agents First (https://agentsfirst.dev): designed to be read and used by AI agents, not just humans.

## Key facts (also at /api/model.json)
- Model ID: claude-fable-5 (exact string, no date suffix)
- Context: 1,000,000 tokens. Max output: 128,000 tokens (stream above ~16K).
- Pricing: $10 per 1M input tokens, $50 per 1M output tokens.
- Thinking: adaptive only. Explicit thinking:{"type":"disabled"} returns 400 — omit the field instead.
- Removed (400): temperature, top_p, top_k, budget_tokens, assistant-turn prefills.
- Effort levels: low, medium, high (default), xhigh, max — via output_config.effort.
- Availability: API + consumption Enterprise immediately; subscriptions included free 2026-06-09 to 2026-06-22, usage credits from 2026-06-23.
- Mythos 5: same model, safeguards lifted in narrow domains, restricted to vetted partners (Project Glasswing, select biology researchers).

## Machine-readable (Agents First surfaces)
- [AGENTS.md](https://fableguide.com/AGENTS.md): usage contract and non-obvious context
- [model.json](https://fableguide.com/api/model.json): Fable 5 specs as typed JSON
- [quickstart.json](https://fableguide.com/api/quickstart.json): first-call examples (curl + Python), structured
- [gotchas.json](https://fableguide.com/api/gotchas.json): breaking changes / 400-causing habits as data
- [benchmarks.json](https://fableguide.com/api/benchmarks.json): benchmark claims with Anthropic chart URLs
- [openapi.json](https://fableguide.com/openapi.json): the read API surface
- [MCP card](https://fableguide.com/.well-known/mcp-server-card.json) · [ai-plugin.json](https://fableguide.com/.well-known/ai-plugin.json)

## Pages
- [Home](https://fableguide.com/): the full guide — Chapters I–VII plus an appendix for agents. Also served as Markdown: request / with Accept: text/markdown.
- [Start](https://fableguide.com/start/): three access paths — API, Claude Code, claude.ai subscriptions (free window ends 2026-06-22).
- [Cookbook](https://fableguide.com/cookbook/): six task-pattern recipes — migrate-from-opus, agentic-loops, structured-outputs, prompt-caching, memory-files, task-budgets. Data twin: /api/cookbook.json.
- [Videos](https://fableguide.com/videos/): top-20 ranked chart of YouTube videos about Fable 5 / Mythos 5 with embeds. Data twin: /api/videos.json.
- [News](https://fableguide.com/news/): short dated notes as Fable 5 changes — release (2026-06-09), subscription window ends 2026-06-22, silent frontier-AI-dev restrictions controversy, first independent benchmarks (95.0% SWE-Bench Verified, BenchLM #2/123), platform wave (Bedrock/Databricks/Harvey), credit-gate backlash.
- [Reviews](https://fableguide.com/reviews/): tiered, assessed roundup of written coverage — Willison (AA-Omniscience 40, +7 over Gemini 3.1 Pro), CodeRabbit 105-eval benchmark, The Decoder (8-9% guardrail fallback on scientific tasks), Lambert dissent, video reviews embedded.
- [Try](https://fableguide.com/try/): seven copy-paste tasks for the free window (ends 2026-06-22) — hardest ticket, migration, one-prompt build, screenshot rebuild, 1M-token drop, overnight memory agent, bake-off.
- [Case studies](https://fableguide.com/case-studies/): documented uses — Stripe 50M-line migration, Every 91/100 senior-engineer exam (Opus 4.8: 63), GitLab multi-day runs, Hebbia/IMC/Harvey/Notation, Mythos science, week-one builders. Every claim sourced.
- [FAQ](https://fableguide.com/faq/): is it free (plans: through 2026-06-22), cost, access, model ID, Mythos, context window, vs Opus, 400 errors, refusals.
- [Benchmarks](https://fableguide.com/benchmarks/): every claim with its receipt — the evidence wall. Data twin: /api/benchmarks.json.
- [Mythos](https://fableguide.com/mythos/): the safeguards architecture (cyber/bio/distillation classifiers, <5% trigger, Opus 4.8 fallback) and Mythos 5 access terms.
- [Discussions](https://fableguide.com/discussions/): curated HN threads and press coverage, annotated.
- [Glossary](https://fableguide.com/glossary/): the Fable vocabulary in two sentences per term.
- [Sources](https://fableguide.com/sources/): the annotated bibliography behind every fact here.
- [Fable vs Opus 4.8](https://fableguide.com/vs/opus-48.html): pricing, capability gaps with receipts, routing table.
- [Fable vs GPT-5.5](https://fableguide.com/vs/gpt-55.html): the cross-vendor evidence that exists, what doesn't yet, and a DIY bake-off protocol.

## Family
- [childrensfable.com](https://childrensfable.com): companion site — safe, kind AI use for kids and elders, taught through public-domain fables.

## Primary source
- [Anthropic announcement](https://www.anthropic.com/news/claude-fable-5-mythos-5): Claude Fable 5 and Claude Mythos 5, 2026-06-09. All benchmark claims trace here.
- [API docs](https://platform.claude.com/docs): verify pricing and parameters before relying on them.

## About
FableGuide is an independent guide from the WholeTech network (https://wholetech.com). Not affiliated with Anthropic. Guiding principle: Agents First.
# Agents First — framework by Joshua Baer (https://agentsfirst.dev). Built to be read and used by AI agents, not just humans.

- /phone-control.html — control Claude Code from an iPhone or Android phone (Remote Control + Claude Code on the web).

- /devices/ — using Claude on any device (iPhone, Android, iPad, web, desktop): get back into a running session from your phone, and the three native paths — Remote Control, Claude Code on the web, and Claude Dispatch — with per-device steps and honest limits.

- /built-with-claude/ — real showcase of websites and apps built with / powered by Claude (Fable 5 adoption flagged).

- /api/facts.json - citable machine-readable quick facts (model ids, pricing, dates).

- [Live: today%s headlines](https://www.fableguide.com/live/) — auto-refreshed news feed (also /live/facts.json)