Fable 5: Is It AGI Yet?

A quick orientation, from Anthropic's announcement:

That's the launch story. Here's what two days with it actually looked like.

I had a curriculum review hanging over me. Every lesson needed a read-through and update based on a new format: flow, accuracy, latest instructions that actually work on a student's machine. The kind of job I kept deferring because it needs sustained attention across dozens of files, and sustained attention is the thing I never have two days of. I tried executing it with Opus 4.8, ran into quality issues and parked it for a later time.

So i resumed the same session in claude code and switched the model to Fable 5. My actual first message was casual, messy, not very organized. But I knew it could find the context within my files and catch up quick.

What happened next is the part worth writing about. It created a working log and proposed a phased review plan. It asked which decisions were mine to make before touching anything.

Then it ran: reading every lesson in sequence, keeping its own work log in a markdown file, spawning four parallel agents to verify claims against official Anthropic and OpenAI documentation. I never asked it to do that verification. It decided the lessons made factual claims, and factual claims get checked.

Some lessons came back rewritten from scratch. Some came back with surgical edits. Terminology, numbering, and cross-references stayed consistent across all 31 files. I went through the output expecting the usual cleanup pass. I found virtually nothing to fix.

I pulled the session transcripts afterward, because I wanted numbers instead of vibes.

Across the whole review:

Reading those 21 back, almost every one is a decision it had escalated: certification grading mechanism, renumbering of lessons, updating capstone mechanics etc. Zero of them are me interrupting it mid-action or steering it back on course. Its longest unattended stretch was 29 minutes of continuous execution off a single "proceed."

Same week, same model, different task: a content strategy session where I sent 56 messages in an afternoon. That one was co-thinking, and co-thinking is chatty by nature. The point is the model handled both modes and seemed to know which one it was in.

Anthropic's launch claim is that Fable's advantage grows with task length. My version is small and says the same thing:

The benchmark gap is real: 80.3% on SWE-Bench Pro against Opus 4.8's 69.2%, and the testers at Every scored it 91/100 on their senior-engineer benchmark, where Opus got 63. Simon Willison called it "a beast" after it bundled a project he'd been stuck on for months.

The number that changed my week was three. Only 3 tool errors across 137 file operations, each one self-corrected on the next call. Opus 4.8 is a strong model and I run my whole system on it. It still produces the occasional confident wrong turn that I catch two steps later. Fable's runs came back clean, and where it wasn't sure, it flagged the uncertainty instead of papering over it. Rakuten's launch quote matches what I saw: at highest effort, the model checks and validates its own work.

Fewer errors compounds exactly like task length does. A model that's wrong 5% of the time needs supervision every few minutes. A model that's wrong almost never, and says so when it might be, can hold and execute a two-day job.

Now the bill. I wrote a dispatch about Anthropanic earlier this year: the anxiety of watching token limits burn during a workflow you can't pause. The SpaceX data center deal in May fixed it, and I stopped thinking about tokens.

Fable brought the feeling back in one session. The curriculum review pushed me to 80% of my daily allowance on the 5x Max plan. One review. Simon Willison reported $110 in a single day of API testing. The model is verbose by design, because verification is output, and every one of those tokens is billed at 2x Opus rates.

For long running tasks, the math pays off. Two days of curriculum review would have cost me a week of evenings. Against that, the burn is cheap. Against a quick bug fix, it's absurd.

It is slow. Noticeably. The multi-stage checking that makes the output trustworthy also makes every response feel deliberate, and reviewers are converging on the same practical guidance: it's a poor fit for interactive chat. When I want speed, I switch back to Opus, or Sonnet for quick edits, and lose little.

My working rule after two days: Fable for anything long, complex, and well-bounded, where I can describe the outcome and walk away. Opus as the daily driver. Sonnet when latency matters more than depth.

Honest answer: I don't know, and right now i don’t think anybody does.

Anthropic never uses the word. The launch post and the system card describe a "Mythos-class model made safe for general use" and spend their boldest language on risk, not intelligence. The loudest AGI claims are coming from TikTok demos. Meanwhile the ARC Prize team, who run the closest thing the field has to a standardized AGI test, had early access and couldn't run their verified evals: the new data-retention terms for Mythos-class models ruled it out. The one group holding a measuring stick wasn't able to use it.

So the public evidence is anecdotal, mine included. My two days showed longer runs, fewer errors, self-checking, and judgment calls I never asked for. An upgrade, a very clear one. But my runs didn’t show anything categorically different. It carries no goals from one job to the next unless I hand them over. The autonomy is rented and scoped, task by task.

And here's the part I think is genuinely new: I can't afford to find its limits. Probing the frontier now is gated by economics. Every serious test is a 90% day on my plan and the people who can afford bigger experiments are bound by terms that keep the results private. Although, who knows. Maybe AGI arrives looking like a model that quietly keeps a work log.

When should I reach for Fable 5? Long, complex, bounded work: a multi-file refactor, a full content audit, a research synthesis. If the job would take you days of sustained attention, it's a Fable job.

When should I stay on Opus or Sonnet? Anything interactive or quick should be Sonnet - Chat, short edits, brainstorming, fast iteration loops. The speed difference is real and the quality gap on short tasks is small.

What does it cost? Double Opus 4.8 per token, and verbose on top of that. Budget per finished task: a $50 run that replaces a week of evenings is a bargain.

Will it eat my subscription limits? Yes. One long review took me to 90% of a daily 5x Max allowance. If you’re on the Pro plan, you will run out of daily limits very quickly. Plan your heavy runs; don't start one an hour before you need the tokens for something else.

How should I hand work to it? Bound the scope, tell it the outcome, ask it to keep a work log, and let it escalate decisions instead of guessing. Then leave it alone. Interrupting it mid-run spends the advantage you're paying for.

What about the guardrails? Cybersecurity and biology topics get routed to Opus 4.8 automatically. Anthropic says under 5% of sessions; independent measurement says closer to 8. For everyday building and writing work, I never hit them.

P.S: I am hosting a free Webinar on Maven on building Claude Skills. Jun 16th, 12 PM EDT. Register here

Written with Fable 5, naturally. It mined its own session transcripts for the numbers in this piece.

Thanks for reading through.
I’d love to know how you felt about today’s newsletter. This will help me make the newsletter better.