how to actually use Fable 5.

In today’s email:

• when Fable is worth 2x the price (and when it isn’t),

• the 3-step kickoff prompt workflow for hard projects,

• a 2-minute CLAUDE.md upgrade,

• and 7 reads worth your time.

Hey,

okay, quick recap, because this was a wild few weeks:

• June 9: Anthropic launched Claude Fable 5, their biggest model ever. A new tier above Opus.

• June 12: the US government export-controlled it. A researcher got it hunting software vulnerabilities a little too well, and since Anthropic couldn’t verify anyone’s nationality in real time, they pulled it. Worldwide. Three days after launch.

• July 1 (yesterday): controls lifted. Fable 5 is back globally, with new safety classifiers.

Here’s the part that matters for you: Fable 5 is included in your Claude Pro/Max plan only until July 7. After that it moves to pay-per-use credits at API rates: $10 in / $50 out per million tokens, double the price of Opus.

So this week is the free trial. And most people are going to waste it asking Fable to fix typos.

Let’s not be most people.

Give it an everyday task, and it’s just an expensive, slower Opus. This isn’t my opinion. One team ran hundreds of simple tasks through both: identical accuracy, ~1.3x slower, 2x the price. The Hacker News verdict on ordinary coding was “mid-tier results,” and they’re not wrong.

Give it a hard, well-specified task, and it’s a different species:

• It one-shotted a compiler bug that Opus failed 16 consecutive times.

• Every’s “senior engineer” benchmark: Fable 91/100. Opus 63. GPT-5.5 62.

• One audit found it cost $4.40 per real bug found vs $17.55 with Opus. Cheaper per outcome despite double the token price, because it one-shots what other models retry.

When to reach for it:

• The bug or refactor you’ve been stuck on for weeks

• One big end-to-end build from a clear spec (people are running it 10+ hours unsupervised with /goal in Claude Code)

• Anything where being right the first time is worth more than being fast

When not to:

• Quick questions, small edits, everyday coding. Opus or Sonnet: same accuracy, half the price.

• Anything security-flavored. The new classifiers will refuse and hand you to Opus anyway. And if you keep requesting it to do things you’re not allowed to do, expect to be severely limited on access.

• Anything you want fast. 15-minute turns are normal on hard tasks. That’s not lag. That’s the product.

Anthropic’s own prompting docs say it, and the community confirms it: the models changed, and most people’s habits didn’t.

• Describe the outcome, not the steps. Step-by-step prompts actively make Fable worse. Manage it like a new hire, not a chatbot.

• Whole spec in ONE message. Drip-feeding instructions across turns measurably hurts performance.

• Don’t babysit it. It verifies its own work. Set it running, check back.

• Don’t run max effort. It overthinks. One test drained the daily allowance of a $100 Max plan in under 9 minutes. Default high is the sweet spot.

Notice what all four have in common:

The skill isn’t prompting anymore. It’s spec-writing.

Which brings me to the thing I actually want you to steal this week.

A three-step workflow that matches how this model actually works:

Step 1: Write the PRD with a cheap model. Open Opus 4.8 and say: “Interview me one question at a time to fill in this template.”

1. What we’re building (one sentence)

2. Who it’s for + the problem (Fable performs better when it knows why)

3. What “done” looks like: checkable criteria, not vibes

4. Core user flow (3-7 steps)

5. Screens (aim for 5 or fewer)

6. What data gets stored (plain language)

7. Stack + integrations

8. Out of scope: what NOT to build

9. How it should verify its own work

Step 2: Kick off Fable, end to end. /model fable in Claude Code (or pick it on claude.ai), paste the entire PRD as one message, set /goal, walk away. No follow-up nudges.

Step 3: Review and iterate on a cheaper model. Switch back to Opus or Sonnet for tweaks. Fable did the expensive thinking; don’t pay Fable prices for polish.

Spec-writing is cheap. The build is one long run. That’s the whole economics of this model in three steps.

And if you think this is only for real engineers: a self-described non-developer used Fable to build a full newsletter analytics dashboard while on vacation. Someone else shipped a browser MMORPG (classes, dungeons, parties) in two days. A Minecraft clone took 37 minutes and $12 of API credit.

The window closes July 7. Pick your hardest project and run the Kickoff this week.

• Sonnet 5 intro pricing: worth it. Launched quietly alongside Fable’s return at $2/$10 per million tokens through August 31. For 80% of your daily work, this is the value play right now.

• awesome-claude-fable-5: a running list of nearly 90 real Fable builds with sources. Good for calibrating what “hard enough for Fable” looks like.

• World of ClaudeCraft: that two-day MMORPG, MIT licensed. Skim the repo to see what a well-run long Fable session actually produces.

Open your CLAUDE.md (or system prompt) and delete every line like “CRITICAL: double-check your work” and “you MUST verify before finishing.”

That scaffolding was written for weaker models. The current generation follows instructions literally, and over-instruction now degrades output. Anthropic says this in their own docs. Two minutes of deleting is the cheapest model upgrade you’ll get this year. And token-efficiency and caching will be improved, too.

• What it feels like to work with Mythos-class AI: Ethan Mollick’s write-up is the best “what this actually feels like” piece so far.

• Simon Willison’s Fable 5 notes: including burning through $110 worth of tokens in one day, on a $100/month plan.

• The Fable 5 vibe check: Every’s benchmark writeup. Their verdict: “a warp drive for power users—but overpowered for everyone else.”

• The orchestrator playbook: Fable plans, Sonnet/Haiku execute. Same audit, 74% cheaper. This is the pattern to copy.

• The Fable 5 deep-dive guide: the most rigorous independent testing we found, including where the launch hype didn’t survive a retest.

• Fable refuses to say “Hello”: The Register on launch-week classifier false positives. Funny, and a fair warning if you do security work.

• Fable 5 prompt library: copy-paste starting points, sorted by use case.

One ask before you go: hit reply and tell me the hardest thing you’d point Fable 5 at, the project you’ve been circling for months. I read every reply, and I’ll feature the best builds in an upcoming issue.

Thanks!