Goals, workflows, learning paths & demos — a practical guide to Anthropic's first public Mythos-class model.
Claude Fable 5 is Anthropic's most capable widely released model, launched on 9 June 2026. It shares the same underlying model as Claude Mythos 5 — Anthropic's restricted frontier model, available only to approved organisations through Project Glasswing — but adds a layer of safety classifiers that make it suitable for broad release. The practical difference from earlier Claude models is the shape of work it can carry: Fable 5 is built for long-running, asynchronous execution. It plans across stages, calls tools, reads results, delegates to subagents, and checks its own work. It is less a chat partner and more a delegate you brief with a goal and verification criteria.
Fable 5 and Mythos 5 share one underlying model. Fable 5 carries safety classifiers; Mythos 5 (without them) is limited to vetted organisations — cybersecurity teams via Project Glasswing and approved life-sciences researchers.
Classifiers target offensive cyber techniques, sensitive biology/chemistry, and extraction of the model's reasoning. Flagged queries are answered by Claude Opus 4.8 instead. Anthropic expects triggers in under 5% of sessions on average.
The classifier layer is what makes broad release possible at all. Early reports note false positives near the boundaries (e.g. benign biomedical queries), so plan for fallback behaviour in any workflow that touches those domains.
Fable 5's economics and capability profile push teams toward an escalation model: cheaper models handle volume, Fable 5 handles the reasoning step where quality changes the outcome. These are the figures to keep in view.
For most day-to-day work, you should. Early consensus is that routine tasks undersell Fable 5 — output on simple work looks much like Opus 4.8 at twice the cost. The model earns its price on long-horizon, ambiguous, high-leverage tasks: large migrations, multi-stage agent workflows, dense document analysis, and vision-heavy work. Anthropic's own guidance is that teams seeing the best outcomes apply it to their hardest unsolved problems. The right question isn't “which model is stronger?” but “which model gives the best cost per successful task?”
Anthropic's official Fable 5 prompting guide marks a shift: the prompt's job moves from steering each step to setting direction and checking results. Habits that were best practice a generation ago — step-by-step scaffolding, chained messages, over-specified instructions — can now work against you. Six patterns matter most.
Give the goal, the context, the constraints, and what “done” looks like. Anthropic's guidance: don't test it only on simple workloads — its capability range shows on work that previously took hours to weeks. Inside Anthropic, the Claude Code team reports the human role shifting from checking whether Claude is doing the work correctly to steering whether it is doing the right work.
// Instead of 12 step-by-step messages: Goal: Migrate the billing service off v1 API Context: repo, constraints, owners, deadlines Done when: tests pass + diff reviewed + report
At higher effort settings Fable 5 reflects on and validates its own output before returning it. Self-verification works best when there's something concrete to check: tests, a schema, a reference output, or explicit acceptance criteria. Defining the evaluation before delegating the work is the single highest-leverage habit.
Fable 5 exposes effort levels. High effort means deeper planning, checking, and context expansion — and slower time-to-first-token, since heavy reasoning runs before the answer. Counterintuitively, capability and speed aren't always opposed: on launch spreadsheet benchmarks it beat Opus 4.8 at every effort level while finishing 25–30% faster. Match effort to stakes, not habit.
Fable 5 dispatches parallel subagents more readily than prior models. Anthropic recommends giving explicit guidance on when delegation is appropriate, preferring asynchronous communication over blocking on each subagent, and keeping long-lived subagents that retain context — which saves cost through cache reads.
Old prompt habits like “show your thinking” or “repeat your internal reasoning” can trigger the reasoning_extraction refusal category — strip them from migrated prompts, skills, and system instructions. Benign security and life-sciences work mostly passes, but work near those boundaries should plan for the Opus 4.8 fallback or use a different model.
The official guide notes a behavioural shift: when a user is describing a problem or thinking out loud, the deliverable is the assessment — Fable 5 is steered to report findings and stop, not apply fixes unasked, and to check evidence before state-changing commands. Structure your harness and review flow around that.
Fable 5 is available in the Claude apps, Claude Code and agent harnesses, the Claude API (model string claude-fable-5), and cloud marketplaces including Amazon Bedrock and Claude Platform on AWS. Wherever it runs, the workflow pattern that's emerging is the same: route by task complexity.
Document-heavy analysis, contract review, complex spreadsheet and finance work, long-form research synthesis, and vision tasks — reading charts, PDFs, UI screenshots, and scientific figures with precision.
Codebase-scale migrations, multi-file refactors, and multi-hour agentic sessions. Caution: agentic tasks can silently branch into 4–6 concurrent subagent calls, multiplying token consumption — set spending caps and watch usage-credit overflow defaults.
Production deployments treat Fable 5 as a high-capability reasoning engine inside a governed architecture: model routing through a gateway, prompt caching for the 90% input discount, refusal/fallback handling, and cost telemetry per task.
| Task profile | Route to | Why |
|---|---|---|
| Long-horizon, ambiguous, high-leverage — migrations, deep research, multi-stage agents, vision-heavy documents | Fable 5 | Largest capability lead on long, complex tasks; plans, delegates, and self-verifies. Worth the 2× price when a correct answer saves hours of senior work. |
| Standard agentic coding and knowledge work, day-to-day volume | Opus 4.8 | Half the price, strongest generally available cost-to-quality; also the model that answers Fable 5's safeguarded topics. |
| Routine, latency-sensitive, or high-volume work — rewriting, classification, short drafts | Sonnet / Haiku | Fable 5's token appetite and output pricing make small models the rational default for volume; reserve premium reasoning for the steps that need it. |
| Work near classifier boundaries — security tooling, life-sciences methods | Opus 4.8 / Mythos 5* | Fable 5 falls back on these domains; documented false positives at the edges. *Mythos 5 requires approved access (Project Glasswing / vetted research). |
Launch demos from Anthropic and early community builds give the clearest picture of the capability range. Treat them as evidence of first-draft quality, not finished-product proof — the consistent lesson is that ambitious first drafts got dramatically easier, while taste, constraints, and review still belong to you.
Built a solar-system simulation deriving planetary orbital motion from physics first principles, then used it to predict solar eclipses. Shown on the official launch page.
Plays Factorio autonomously — strategising and building an automated factory — and navigates Pokémon FireRed from raw screenshots. A signal of sequential reasoning and spatial understanding, not a party trick.
Designed a complete 3D-printable model inside a browser-based CAD editor that Fable 5 itself created — including the editor's built-in AI copilot. Also: a fluid simulation synchronised to music it composed in code.
Developers shipped Minecraft-style interactive 3D worlds, a Three.js recreation of Yosemite Valley with terrain data and procedural forests, a 3D Library of Babel explorer, and one-shot playable games like a self-aware Snake.
Early builds include a Lovable-style app builder, a music studio, an all-in-one productivity app, and a markdown editor — first drafts with deeper interface detail and working mechanics, needing fewer follow-up prompts.
A GitHub repo curating 60 sourced cases — coding agents, long-running automation, games, 3D simulation, knowledge work, benchmarks, and launch limitations — each with original source, creator attribution, and evidence type. The best single index of reproducible examples.
Anthropic Academy is Anthropic's official learning platform — courses hosted on Skilljar at anthropic.skilljar.com and indexed at anthropic.com/learn. The 2026 catalogue spans 13+ self-paced courses, all free, each awarding a certificate on completion. Only an email is needed to enrol. Courses fall into three tracks.
Co-developed with academic partners (Prof. Joseph Feller, UCC; Prof. Rick Dakan, Ringling) around the 4E framework — Effective, Efficient, Ethical, Safe. Variants include Framework & Foundations, Teaching AI Fluency, editions for students and educators, a nonprofit edition with GivingTuesday, and AI Capabilities and Limitations.
Claude 101 covers everyday work tasks and core features — no coding, no Anthropic account required. Claude Code 101 introduces the agentic coding loop (Explore → Plan → Code → Commit), and Introduction to Claude Cowork covers the agentic desktop app for knowledge work.
Building with the Claude API, Claude Code in Action, Introduction to MCP (servers and clients in Python; tools, resources, prompts), MCP Advanced Topics (sampling, notifications, transports), Introduction to Agent Skills, and Introduction to Subagents.
| You are | Path | Outcome |
|---|---|---|
| New to Claude entirely | Claude 101 → AI Fluency | Productive everyday use plus a framework for when (and when not) to delegate to AI. |
| A builder heading to production | Claude 101 → Claude API → Claude Code in Action → Intro to MCP → MCP Advanced → Agent Skills | Zero to production Claude applications with MCP integrations and Skills. |
| An agent-workflow operator | Claude Code 101 → Claude Code in Action → Agent Skills → Subagents | The densest developer path: the agentic loop, operational workflow, reusable skills, then delegation and parallelism — the exact skills Fable 5's subagent behaviour rewards. |
| Fable 5 specifically | Official prompting guide + Intro page | Anthropic's model-specific docs cover effort levels, instruction following, long-run progress claims, memory systems, and the reasoning_extraction category — the migration checklist for existing prompts and harnesses. |
Academy materials are currently English-only (community-reported Spanish subtitles exist for Claude Code in Action). Fable 5 itself launched after most courses were published — pair the Academy paths with the model-specific docs above for current behaviour.
It is not the default model — it's the escalation tier. The honest test: does the task have a long horizon, real ambiguity, large context, agentic execution, or high leverage? One or more of those, and Fable 5 starts paying for itself.
The task is long-horizon: many steps, decisions, validations, revisions — large code migrations, multi-file refactors, multi-stage agent workflows.
The work is document- or vision-heavy: dense PDFs, charts, scientific figures, UI screenshots, contract review, complex finance analysis.
A correct answer saves hours of senior work — the cost-per-successful-task maths favours premium reasoning.
You can define verification before delegating: tests, schemas, acceptance criteria the model can check itself against.
The work is routine — short rewriting, brainstorming, translation, classification. Output looks like Opus 4.8 at twice the price.
Latency matters: heavy pre-answer reasoning makes it slow to first token for interactive use.
You're near classifier boundaries — security tooling or life-sciences methods — where fallback and documented false positives will disrupt the workflow.
Costs are uncapped: agentic sessions with parallel subagents have produced reported $100 burns in single sessions. Set hard spending caps before, not after.
Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost from 9–22 June 2026 (counting as 2× usage). From 23 June it is removed from those plan limits and continued use requires usage credits billed at API rates. Anthropic says it aims to restore it as a standard plan feature once capacity allows. If your workflow depends on frequent Fable 5 sessions, use the included window to test your hardest tasks and audit whether the maths works at usage-credit pricing.
Model string claude-fable-5. The 5:1 output-to-input ratio matters: long reasoning, code generation, and agent reports produce large outputs, and tool schemas and tool results all count as tokens. Prompt caching cuts input costs by 90%.
Once a plan's usage window is exhausted, Claude can continue drawing on pay-as-you-go usage credits — by default charged automatically. Check whether your account has a usage-credit spending limit configured before running long agentic sessions.
Available on Amazon Bedrock and through the Claude Platform on AWS for teams that need to build inside an existing cloud environment with consumption-based Enterprise billing.
Launch announcement — capabilities, safeguards, demos ↗
Introducing Fable 5 & Mythos 5 — API, billing ↗
Prompting Claude Fable 5 — model-specific guide ↗
Prompting best practices — cross-model reference ↗
Claude.ai Help Center — plan limits and product behaviour ↗
Anthropic Academy — official learning hub ↗
Academy course catalogue — all free, with certificates ↗
Claude 101 — the standard starting point ↗
awesome-claude-fable-5 — 60 sourced community cases ↗
Anthropic launch post · Anthropic prompting guide · AWS News Blog · CNBC · TechCrunch · MacRumors · The Register · Anthropic Academy / Skilljar · awesome-claude-fable-5 (GitHub) · Developers Digest (usage limits)
Content validated June 2026. Pricing and plan availability change quickly — verify against Anthropic's docs before relying on figures. Claude, Claude Fable, and Claude Mythos are trademarks of Anthropic PBC. This is an independent educational guide by Imbila.AI.
Fable 5 rewards delegation, description, discernment, and diligence. Our interactive Claude Academy course teaches exactly that — free, self-paced, with an AI tutor.