Claude Fable 5: you're the bottleneck

By Jeroen Vonk — 8 July 2026

In late June, Thariq Shihipar took the stage at the AI Engineer World’s Fair in San Francisco. Shihipar works on the Claude Code team at Anthropic. His talk was called “A Field Guide to Fable”, a guide to the strongest Claude model you can use right now. One line stuck with me. It’s about you, not the model.

Watch Thariq Shihipar’s talk on YouTube: A Field Guide to Fable.

What Claude Fable 5 actually is

Anthropic released Claude Fable 5 on 9 June 2026. It’s the first model from the Mythos class made available to the general public, a tier that sits above Opus. Under the hood it runs the same core model as Claude Mythos 5, with a few extra safety layers on top so it’s safe enough for general use.

In plain terms: this is the most capable Claude model anyone can switch on. It leads on software engineering, knowledge work, reading images, and scientific research. The best-known example comes from Stripe. The company had Fable 5 convert fifty million lines of Ruby code to a newer version in a single day. A team of engineers had spent more than two months on the same job.

It comes at a price. Fable 5 costs ten dollars per million input tokens and fifty dollars per million output tokens. That’s twice as expensive as Opus 4.8. The model was briefly pulled over US export controls, but it’s available worldwide again. If you want to call it directly, the model string is claude-fable-5.

The core message: you’re the constraint, not the model

Shihipar put it this way:

“The binding constraint in agentic AI is no longer the model. It’s you: the clarity of your own thinking before the task begins.”

For years the model was the limiting factor. You gave it a task, and the question was whether the AI could handle it. Often it couldn’t. So you broke the work into small pieces, checked every step, and caught the spots where the model stopped too early.

With Fable 5 that flips. The model works on its own for hours, tests its own output, and sometimes delivers better results than you would have. The question isn’t whether it can do the job. The question is whether you described the job well. The clearer you are about what you want before you start, the better what comes out. Vague request in, vague result out. That’s the deciding factor now, and it sits with you.

Three practical lessons from the keynote

It’s a field guide, so there are concrete lessons you can use today.

1. Think first, prompt second. The temptation is to type a question straight into the chat box. Don’t. Take a minute first to work out what you actually want. What does done look like? Who is it for? Shihipar sometimes has the model ask him questions about a task before it starts. That forces you to think and fills the gaps in your own brief.

2. Give context, not bare rules. Don’t say “keep it simple”. Say why. For example: “this is an experiment, there’s a real chance we throw it away in a month, so don’t build anything that’ll be painful to delete.” The model makes better choices when it understands why something matters, not just that it does. Context beats instructions.

3. Give a goal plus a way to check it. Instead of scripting every step, hand the model a goal and tell it to review its own work. Ask for a short report: what got built, and where it deviated from the plan. That’s how your role shifts from watching to steering.

When to use Fable 5, Sonnet 5, or Opus 4.8

Fable 5 isn’t always the right call. It’s expensive and heavy, and for a lot of work you don’t need that power. A rough guide:

Model	Strong at	Cost	When
Sonnet 5	Fast, everyday tasks, high volume	Low	Email, summaries, standard copy, first drafts
Opus 4.8	Heavy thinking at half the Fable price	Mid	Tricky analysis, code, planning, most serious work
Fable 5	The biggest, longest, most complex jobs	High	Rebuilding a whole app, large migration, deep research

In practice: start with Sonnet 5 or Opus 4.8. Reach for Fable 5 only when a job is so large or complex that the extra power is worth the price difference. For a freelancer who needs the odd bit of text or a calculation, Fable 5 is overkill.

What this means for a small business

You don’t need to be a developer to get something out of this. The point of the keynote applies to any work you hand off to AI.

Say you do the books and you want a recurring report drawn up. The old reflex: explain step by step what the model should do. The new approach: describe what the report needs to show, who it’s for, what decision the reader makes from it, and let the model fill in the rest. Then ask it to check that every figure matches the source.

Or you run a webshop and you want to revise your product copy. Not “make this text better”, but “this is for hesitant buyers who think the price is steep, show them why it’s worth it, in my own plain tone.” The model could write that copy all along. Whether it turns out well depends on how well you described the job.

That’s the real work no model release takes over. The tools get so good that the value shifts to the person who gives the right instruction. That’s where your edge is, and it’s also where you lose it if you’re sloppy.