First, let's clear up the confusion: Fable or Mythos?

I mixed these two up at first, and you probably will too, so let's settle it upfront.

Back in April, Anthropic introduced a model called Mythos. It was so good at finding security vulnerabilities in software that they didn't release it publicly, fearing it could fall into the wrong hands. They only gave it to a handful of selected companies through a cybersecurity initiative called Project Glasswing.

Fable 5 is the public release of that Mythos-class model. Same underlying model, with a safety layer on top that catches and blocks high-risk topics. So it's not a "watered-down, weaker model." Quite the opposite: it's the most capable model Anthropic has made generally available so far. The brakes are only applied in a narrow set of potentially dangerous areas like cybersecurity and biology.

Getting this distinction right matters, because the question "is this Mythos-level or something restricted?" can create unnecessary disappointment. The answer is clear: for coding, knowledge work, and vision, basically everything you do, you're using the full capability.

Where does it make the biggest difference?

In one sentence: long-horizon, autonomous agentic coding.

It's good at raw code generation too, but that's not where the leap is. The difference is the model's ability to sustain a task on its own for hours, even days. It understands an intent, plans the steps, calls tools, checks its own work, and keeps the task moving without constant input from you.

The most striking example came from Stripe: on a 50-million-line Ruby codebase, it completed a sweeping migration that would have taken a team more than two months, in a single day. Let me add a parenthesis from my own experience here. When I previously tried handing a large Angular version migration to Claude, the problem was never that the model couldn't fix a single file. The problem was that by the time it reached file 15, it had forgotten the decision it made in file 3. Fable 5's whole claim sits exactly at that point: finishing the long job without losing context.

What does this do for me on the frontend?

I see three concrete areas, and I plan to test all of them on Cinekholm one by one.

1. Large migrations and complex refactors

Angular major version upgrades, big dependency bumps, monorepo refactors. Not single-file edits, but an entire migration complete with planning and tests. Cinekholm still has a few modules written with older patterns; moving them to a Signal-based structure is exactly the kind of multi-file, context-heavy work this is built for.

Here's the prompt skeleton I'm planning to use:

Look at the src/app/features/catalog module in Cinekholm.
Apply the following steps in order, and after each step stop and
give me a summary:

1. Detect all RxJS + async pipe usage, list which ones can be
   converted to signals.
2. Move components to OnPush, clean up unnecessary subscribe() calls.
3. For every change, update the existing test files, or write them
   if missing.
4. After the change, run `ng test catalog` and show me the output.

Don't rewrite the whole module at once; go component by component
and wait for my approval after each one.

The critical part here is "stop after each step." It positions the model as a junior you supervise. It removes the risk of it tearing through the whole module at once and leaving you unable to review the result.

2. Design to code (vision), my Figma workflow

This is the part I'm most excited about, so let me expand on it. My current Figma workflow roughly goes like this: a designer prepares a screen, and I read the dimensions, colors, and spacing values out of Figma one by one and translate them into a component. The tedious part isn't the creative work, it's the mechanical work. Maybe 150 lines of a 200-line component are just "put this div here, set this padding to that value."

This is exactly where Fable 5's vision side comes in. Anthropic's own example is "rebuilding a web app's source code from screenshots alone." The flow I'm going to test on Cinekholm will be:

I give it a high-resolution screenshot of the Figma screen.
Alongside it, I add my existing design tokens (color variables, spacing scale, typography) as context.
I ask the model to build the screen using my existing component library.
Most importantly: I take a screenshot of its output and say "compare this against the original design and fix the differences."

That last step is the real power of vision. The model isn't just writing code, it can visually verify its own output against the target. So I can hand off the "does this match pixel for pixel" loop as well. On a visual-heavy screen like Cinekholm's movie detail page, I want to measure just how much time this saves.

3. Long-running automation flows

Inside an agent harness like Claude Code, it can run for days; it plans across phases, delegates to sub-agents, and writes its own tests. For me, that maps to multi-step, long-horizon work like the Jira automation setups I've built before.

There's an interesting contradiction on the security side

It might feel counterintuitive: the public Fable 5 is not the model you'd reach for to do security work.

Here's the logic: the model's class (Mythos) is exceptional at finding vulnerabilities, which is exactly what made governments nervous. That's precisely why the capability is restricted in the public version. Questions about cybersecurity, biology, and similar topics get routed to the lower-capacity Claude Opus 4.8.

In practice: general "secure code review," reviewing auth logic, and other defensive work are fine with Fable 5. But offensive security research like exploit development is restricted anyway. According to Anthropic, this routing kicks in for less than 5% of sessions, and you'll rarely hit it in frontend work.

In short: position Fable 5 as a coding and automation tool, not a security tool.

The right mental model to do it justice

You won't feel the difference in single-line autocomplete. Opus is already enough there, and cheaper. To do Fable 5 justice, hand off whole, multi-step jobs.

The analogy that makes the most sense to me: think of it not as an "assistant," but as a junior you supervise. Give it the work, come back, review the output. The value comes less from raw code generation and more from this sustained execution.

Access and cost

The model id in the API is claude-fable-5. It's available through the Claude API, Claude Code, Enterprise plans, and AWS, Google Cloud, and Microsoft Foundry. It's included at no extra cost for Pro, Max, Team, and Enterprise subscribers until June 22; after that it moves to a credit-based model. API pricing is $10 input and $50 output per million tokens.

What's next

Over the coming days I'll genuinely test all three of these flows on Cinekholm: first the signal migration of the catalog module, then building the movie detail page from Figma, and finally the automation side. I'll measure where it actually saves time and where it falls short of expectations, and share the results in a separate post.

Because announcement copy is always shiny. The real question is whether this works on a Monday morning inside your own codebase. I'll try it and write about that too.

Sources: Anthropic's official announcement, the AWS Bedrock announcement, and news coverage from June 9, 2026 (CNBC, Reuters, Inc.).