The Fable 5 Playbook · Updated June 11, 2026

Claude Fable 5 and Mythos 5 — the playbook, from 45 creators at once.

Not another summary. We mined 45 creator breakdowns for the parts that matter: how it beats Opus, how to orchestrate it, what to build, and where the experts disagree — every claim linked to the source.

AnthropicAI codingFable vs OpusOrchestration45 breakdowns and counting

See all 45 breakdowns ↓Browse the full library

80.3%

Agentic coding (vs 69.2% Opus 4.8)

2mo → 1day

Stripe, 50M-line codebase

72 days

Of work · 2 people · 16 hrs

~2×

The price of Opus 4.8

~Jun 22

Free access window closes

Breakdowns synthesized

The 60-second version

→Anthropic shipped two models: Mythos 5 (too powerful to release — locked to government and cyber-defense) and Fable 5 (the public version — same brain, safety leash).
→It is the best coding model available — and the most expensive. ~2× Opus, and it can burn $100 in 8 minutes.
→The pros don’t use it for everything. They plan with Fable, build with cheaper models, verify with Fable — the 10/80/10 method below.
→Free in your plan until ~June 22, then metered. The window to push it hard is closing.

People are asking

The questions everyone’s Googling.

Straight answers, pulled from across all 45 breakdowns.

What is Claude Fable 5?

Fable 5 is the public release of Anthropic's new Mythos-class model — the most capable model they've ever shipped. It tops agentic-coding benchmarks at 80.3% and can build full applications from a single prompt. It's the same underlying model as the restricted Mythos 5, wrapped in a safety layer for general use.

What is the difference between Fable 5 and Mythos 5?

Same weights, different leash. Mythos 5 is the unrestricted model — so capable at finding exploits and designing proteins that Anthropic locked it to government and cyber-defense partners. Fable 5 is what the public gets: identical capability, with classifiers that quietly reroute sensitive topics to Opus.

Is Claude Fable 5 worth it?

For complex one-shot builds, planning, and overnight agentic runs, yes. For everyday work, no — it's roughly 2x the price of Opus 4.8 and slower (no fast mode). The cohort consensus across 42 creators: Opus for daily work, Fable for the hard 20%.

Fable 5 vs Opus 4.8 — which should I use?

Use Fable for planning and one-shotting complex apps; use Opus or Sonnet for execution volume. The phrase that kept coming up is "plan expensive, build cheap." And drop the effort dial before the model — Fable on medium often beats Opus on max.

What can Fable 5 actually build?

A Notion clone in 2 prompts / 45 minutes, a flight simulator with real stall physics, a 767-painting 3D museum, and a working DAW in 15 minutes. It can also rebuild a working web app from a single screenshot.

When does free Claude Fable 5 access end?

Around June 22, 2026. It is included in paid Claude plans until then; after that it moves to metered usage at roughly $10 / $50 per million input/output tokens.

Fable vs Opus, dimensionalized

Better than Opus? Yes — but only where it counts.

Not vibes. Here is what happened in the head-to-head tests creators actually ran.

Flight sim, one-shot

OPUSBuilt a plane that moved left/right but couldn’t fly forward — broken out of the box.

FABLEWorking physics, speedometer, realistic takeoff and stall.

↗ Jono Catliff

Site rebuild, 35-min run

FABLEScraped the real images, nailed the scroll-jacking animation — 35 min, $30.

OPUSBarely moved from round one — 5 min, $2.70. The vision-verification loop is what flipped it.

↗ Leonardo Grigorio

Same 3 apps, side by side

FABLESwept all three — a 767-painting 3D museum and a playable RTS.

OPUSMuseum gallery broken; the RTS was “blobby placeholder geometry.”

↗ Pat Simmons

6 builds vs Opus and GPT-5.5

FABLEPassed a self-verification loop Opus failed; resisted the default purple-gradient AI look.

OTHERSNeeded babysitting; fell into generic styling.

↗ Jay E / RoboNuggets

Reach for FABLE 5 when…

One-shotting a complex, multi-feature app or game
Planning and architecture — the decisions that compound
Cloning from a screenshot (its vision is the unlock)
The instructions are vague and you want it to push back
You want it to verify its own work before returning
Pre-launch security audits (exposed keys, DB config)

Stick with OPUS 4.8 when…

It is a simple landing page or basic UI (same result, half price)
Everyday tasks — “a rocket launcher to squash an ant”
You need speed — Fable has no fast mode (7 min → 30 min)
You are iterating back-and-forth, not shipping a final
Budget is tight — Fable is ~2× Opus / 4× Sonnet

80.3% vs 69.2%

SWE-bench Pro agentic coding

29.3% vs 13.4%

FrontierCode hard problems (~2× Opus)

50M lines → 1 day

Stripe migration (was ~2 months)

directional

All Anthropic-sourced — treat with healthy skepticism

The orchestration method

Plan expensive, build cheap.

The #1 mistake (per Greg Isenberg, Mark Kashef and AI Edge): using the smartest model for everything. The pros treat model and effort as two separate dials. The 10/80/10:

10%

Plan

Fable is the architect. One expensive planning pass → a spec file. Even the best builder can’t fix a bad plan.

Fable 5 · high/max

80%

Build

Type /model mid-session to drop to a cheaper model. It churns the execution volume against the spec.

Opus / Sonnet · sub-agents

10%

Verify

Re-invoke Fable to review, effort scaled to risk. It catches its own bugs. Then ship.

Fable 5 · low–high

Effort beats model. Fable on medium often beats Opus on max — drop the effort dial before you drop the model.

Fable never runs sub-agents. Sub-agents always use cheaper models. Fable conducts; Haiku and Sonnet play.

Less is more. Give it the goal, not step-by-step. Over-stuffing CLAUDE.md makes output worse and pricier.

Stage	Marketing site	3D website	CRM
Plan	Fable high	Fable extra-high / max	Fable max
Build	Opus med/high or Fable med	Opus 4.8 + Sonnet agents	Extra-high workflows
Verify	Fable low	Fable high	Fable high

⚠️ Watch the model selector — Claude Code silently defaults to Fable, and one builder reported a surprise $1,000 bill. Use /usage, keep a hard spend cap, default to low effort. ↗ Mark Kashef ↗ AI Edge

Use-case playbooks

Here’s exactly what to build — and how.

The real methods creators demoed, with the actual tools and the video to watch.

Build a website

↗ Zubair Trabzada · Luke Carter

Higgsfield MCPThree.jsScroll-cinematic SKILL.mdEffort: Medium

A.Cinematic 3D scroll site for ~$1–2. Fable (1M context) orchestrates; Higgsfield MCP generates the hero image and 3D clips, slices them into frames, and pins each to a CSS scroll position. Screenshot an awwwards.com site as your target, then iterate in plain English.

B.Six sites, one prompt each, no reference images. At Effort Medium, Fable builds a clean landing page up to a 3D scroll journey using Three.js (replacing Spline). A Level-6 site came out near-shippable in 25 minutes.

Edit and produce videos

↗ Samin Yasar · Jack Roberts · Danny Why

Hyperframes MCPHiggsfield CLIAuphonic APIManim + MiKTeXRemotion

A.Agentic editor. Hyperframes (HeyGen) exposes a compositing canvas over MCP — Fable adds tweet overlays, teaser clips and chapter cards by voice, non-destructively. Direct it like a human editor.

B.$2/video faceless pipeline. Claude Code is researcher, scriptwriter and assembler; Higgsfield CLI for visuals; Auphonic API for broadcast audio; it word-timestamps and assembles — no timeline editor.

C.3Blue1Brown-style math animations. Claude writes Manim (Python); MiKTeX renders the LaTeX; composite a Higgsfield character in CapCut.

Run a content engine

↗ Greg Isenberg · Jack Roberts

Obsidian1of10 outlier finderAI judge tournament

A.The 4-System engine. Inputs (story, ICP, offer) → Content Brain in Obsidian → weekly research/performance loop → Fable runs it all autonomously for hours.

B.Steal the recipe. Have Fable deconstruct a target channel into an HTML report; find outlier topics with 1of10; feed both back for contrarian research. (Never publish the draft verbatim.)

C.Copywriting tournament. 8 landing-page versions × 5 AI judge personas (CFO, competitor, ICP…) = 40 scores. Kill losers, merge winners. ~$100 vs $50k in wasted media.

Build a full app

↗ Brian Casel · Build Great Products · Nick Puru

ConvexPRD Creator skillSuperconductor/effort extra-high

A.Spec-then-build. Scope in claude.ai → expand to a milestone PRD with the free PRD Creator skill → hand it to Fable in an isolated git worktree. A real Rails feature shipped in ~41 minutes.

B.Interview-first. “You are my technical cofounder. Interview me before writing any code.” 3 rounds of Q&A → full spec → “go.” It self-tested and wrote its own limitations doc.

C.Two-prompt Notion clone. One vague prompt → Fable inferred Electron, built a Convex-backed desktop app, self-screenshotted and self-corrected. 45 minutes.

Automate your business

↗ Brock Mesarich · Samin Yasar

Clay connectorEmail Voice skillEraContextClaude for ChromeFireCrawl

A.Cold outreach engine. Clay finds contacts; an Email Voice skill reads your last 25 sent emails for tone; Fable drafts personalized emails and stages 29 drafts in Gmail.

B.Bookkeeper and subscription killer. EraContext reads your bank data read-only → spend charts; Claude for Chrome navigates the real billing UI and cancels a subscription.

C.Multi-agent research. Spin 4 sub-agents on different sources, make them debate, and surface “the answer that survives all the arguments” as an HTML report.

Make money

↗ Greg Isenberg · Eric Siu

Productized servicesToken cost ≈ $100/job

›Synthetic focus-group firm — 50-variant AI judge tournaments per launch. ~$3k/launch, ~$100 cost.

›48-hour custom software — the interview prompt is the sales call. $5k flat.

›Contract refund firm — Fable reads contracts for auto-renewals and escalators; take 25% of savings found.

›Pre-launch security audits — $10–50k, the use case creators flagged as most underrated.

Steal these

16 things to have Fable do for you today.

Real instructions from real demos. Copy, paste, tweak.

“Interview me as my technical cofounder, then build a local-first business hub for my service company.”

via Nick PuruWatch →

“Build a 3D CAD editor, then design a printable phone stand inside it and export the STL.”

via Nick PuruWatch →

“Build me a macOS Notion clone with a Convex backend — no auth.”

via Build Great ProductsWatch →

“Connect my bank via EraContext, chart my spend, then use Claude for Chrome to cancel a subscription.”

via Samin YasarWatch →

“Spin 4 sub-agents to backtest my stock hypothesis, make them argue, and give me one HTML answer.”

via Samin YasarWatch →

“Find 10 creators in Clay, research each with its own sub-agent, draft a 3-sentence pitch, and have a checker verify it.”

via Samin YasarWatch →

“Write 8 landing-page versions, create 5 AI judge personas, score all 40, and give me the winner.”

via Greg IsenbergWatch →

“Role-play a funded founder who hates my company — here is my P&L, pricing, churn, and last 50 tickets. Attack me.”

via Greg IsenbergWatch →

“Take this 80-page contract — list hidden costs, 18-month regrets, missing protections, and the exact asks.”

via Greg IsenbergWatch →

“Clone this site from a screenshot using Higgsfield for the 3D scroll, then add a color-explosion section.”

via Zubair TrabzadaWatch →

“Build a scroll-driven 3D performance website in Three.js at Effort Medium — no reference images.”

via Luke CarterWatch →

“Deconstruct this YouTube channel and output an HTML report of its title patterns and visual formula.”

via Jack RobertsWatch →

“Install the Manim skill, render a quadratic-formula derivation animation, then add my character.”

via Danny WhyWatch →

“Add the Clay connector, build an Email Voice skill from my 25 sent emails, and stage 29 cold emails in Gmail.”

via Brock MesarichWatch →

“CoWork: turn these CSVs and brand guide into 3 interactive dashboards and a board report.”

via AI ImpactWatch →

“Run a pre-launch security audit on my repo — exposed API keys, insecure DB config, access-control gaps.”

via Jason LeeWatch →

The map

Where 45 creators agreed — and threw hands.

The honest part you only get by watching all of them. Tap any name to watch.

All-in / blown away

“Best model ever shipped.”

Alex Finn BridgeMind Income Stream Surfers Matthew Berman Nick Puru Pat Simmons Luke Carter

Practical / nuanced

“Great — but use it surgically.”

Theo (t3.gg)Matt Wolfe Jono Catliff Brian Casel Samin Yasar Jason Lee Leonardo Grigorio

Contrarian

“You are using it wrong.”

Greg Isenberg David Ondrej Mark Kashef

Skeptical

“Not worth it yet.”

Jan Marshal Zubair (Not For You)AI Tooltip

Cost-critics

“A rich man’s model.”

Ras Mic Riley Brown Lead Gen Jay

Doomers

“It’s over for devs.”

Richard Echols Jesus Martinez

✓ What nearly all 45 agreed on

Best coding model available right now — near-universal.
It is expensive — ~2× Opus, the most-repeated fact in the set.
Reserve it for the hard 20%, hand the rest to a cheaper model.
Treat it like an employee — set a goal, validate the outcome.
The free window closes ~June 22 — then the meter runs.
It torches tokens and overthinks by default — cap it.

✗ Where they threw hands

Worth the 2×? Pat Simmons and BridgeMind say yes; Jan Marshal and Jason Lee say “incremental, not groundbreaking.”
Does it replace devs? Richard Echols: “permanent underclass.” Build Great Products: “your window is now.”
Trust the benchmarks? Chase AI takes them at face value; Matt Wolfe and Mark Kashef call them gamed.
The price and the reroute: sources split on $30 vs $50/M output, and on whether the safety downgrade fires under 5% or ~75% of the time.

The 5 moves that pay

Treat it like an employee.

Hand it responsibilities, not tasks. Point it at a goal and let it loop. It doesn’t clock out at 5.

Remember the bottleneck flipped.

Building is solved. Compete on taste, distribution, and idea quality.

Aim it at ignored data.

Your notes, contracts, churn logs, support tickets. That’s the gold — not one more chat prompt.

Plan expensive, build cheap.

Big brain to think and spec; cheap models for the grunt work.

Make it interview you first.

“Ask me questions before you build.” Force it to push back on your half-baked idea.

The library

All 45 breakdowns, and counting.