Claude Fable 5.2: what's new, what it costs and three one-shot builds

Anthropic released Claude Fable 5.2 and Mythos 5.2 on 9 November. I went through the launch page in a seven-minute video, then gave Fable 5.2 a single prompt and had it make three things: a redesign of my own website, an interactive nuclear reactor and a virtual pet.

This is the written version, with the figures checked. Where I misread a number on camera, the post uses the one from Anthropic's launch table and says so. You'll get who can use which model, what happened to the price, which scores moved, and what the three builds cost me in money and time.

Who gets Claude Fable 5.2, and who gets Mythos 5.2

There are two models in this release. Mythos 5.2 goes to a restricted group: people in Anthropic's cyber verification programmes and in life sciences work. Everyone else gets Fable 5.2, the less restricted version that ordinary users like you and me are allowed to have.

So unless you're in one of those programmes, Fable is the one you'll actually open. One of the headline results below is a Mythos result, and I've marked it, because it doesn't tell you what Fable will do for you.

What the launch benchmarks show

Scores went up on most of the tasks Anthropic reported. These are the ones worth your attention, with my on-camera slips corrected:

• Paid remote work (the Remote Labor Index): 41.3%, more than double Fable 5.1's 17.9% and almost double Opus 5.5's 21.3%.
• Fermat's Last Theorem: Fable 5.2 finished formalising the proof in 30 hours. Anthropic's earlier research model took about 11 days on the same task.
• RoboDojo simulation tests: 64.9% for Fable 5.2, against 39.4% for Opus 5.5 and 27.6% for Fable 5.1. That's 25.5 points over Opus 5.5 and more than double Fable 5.1. I said 24% in the video; 25.5 points is the right gap.
• Cybersecurity (a Mythos result): Mythos 5.2 scores 76.3% against 61.7% for Mythos 5.1, a gain of 14.6 points. On camera I said a 10% gain at half the cost. The gain is 14.6 points, and the launch table shows no cost cut.
• Protein design: the row is de novo binder design. I called it protein folding in the video, which was wrong.

The Fermat result is the one that stuck with me. Going from about 11 days to 30 hours on one long reasoning job is a big change in what you can hand a model and walk away from. Be precise about it, though: Fable 5.2 formalised the proof. It did not solve or prove Fermat's Last Theorem.

My read is that complex reasoning and problem-solving will be where this model gets used most, along with hardware and robotics. The RoboDojo jump is why I think models like this are getting close to useful in commercial settings: manufacturing, industry and, further out, the home.

What Claude Fable 5.2 costs

The per-token price went down. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, a 20% cut from Fable 5.1's $10 and $50.

In the video I said some costs had increased, which makes it sound like Fable got dearer. That's only true in one place: the cost per task at Max effort on Anthropic's chart. Per token, it's cheaper than the model it replaces.

Coding is where the effort setting matters. The coding score went up a meaningful amount, but some settings cost more per task than the previous model did. On the chart, Max cost more than Extra High and scored lower. If you're paying for coding work by the task, don't assume the top setting buys you the best result. On these numbers it doesn't.

Three builds from one prompt

Launch charts only tell you so much, so I wrote one prompt asking for three separate builds and let it run in a single Claude chat. Most of what came back was right in one shot.

1. My website, redesigned

Fable rebuilt the ABAI Solutions site around a new line, "your definitive AI growth partner". The hero is an animation of leads arriving from different places and being filed into a structured CRM. That's close to work I do for clients, like the setup in how a solar installer stopped losing leads to four inboxes. Top to bottom, the page had:

• a call-to-action button under the hero, plus scroll animations and effects
• a closed-deals section where a bottleneck builds up and you can drag it, scale it and clear it (my current site has a version of this)
• results that move on their own, and a set of my case studies
• how I work, as strategy call, blueprint, build, launch, train and improve, then the cycle repeats
• who I am, how I differ from a typical agency, and FAQs
• a closing prompt to find out where your growth is leaking and book a call

2. A nuclear reactor you can control

The second build was there to test visuals. It's an interactive reactor. You move the control rods up and down and watch the power respond: push them further in and power drops, pull them all the way out and you get a runaway reactor. You can click each part to see what it does, zoom into a fuel rod and its tube to watch fission and the bubbles, then go one layer deeper to see what a fission reaction looks like.

For people who learn by seeing, that's the value. The animation changes live as you move the controls, and all of it came from one prompt.

3. A pet in a pocket

The third is a 3D virtual pet, a bit like a Tamagotchi, pitched as small enough to fit in a coat pocket. It spins, it's built entirely in code and it worked first time. You can play with it and swap in a new pet. Fable also made mock-up packaging and a schematic diagram for it. It's a silly use case, but it shows how much product mock-up work one prompt now covers.

What the run cost, and how to repeat it

The whole run came to about $20 over about 45 minutes. It took longer than it needed to, because I'd forgotten to put a couple of things in the prompt. Treat that as one run of one prompt, not a price list. You should be able to rebuild all three for under $20.

The exact prompt is on the free Fable 5.2 guide, with every launch score from Anthropic's table. To run it yourself:

1. 1Make sure you're on a paid Claude plan: Pro, Max, Ultra, Team or Enterprise. Fable 5.2 isn't available on the free one.
2. 2Open one fresh chat and select Fable 5.2 as the model.
3. 3Copy the prompt from the guide. Swap in your own website and anything else you want built before you send it, so you don't pay for the extra time my run took.
4. 4Send it and let it work. If it stops partway, type continue.

Where I'd use it

From what came back, the clearest uses are education, product mock-ups and getting something to market faster. You can iterate on a visual model like the reactor, a product mock-up or a website design, and the result doesn't have the AI slop look of rounded corners everywhere and the other telltale signs.

Keep two limits in mind. Max effort isn't automatically better, as the coding chart shows, and most of the builds were one-shots, not all of them. Budget for a follow-up message or two.

If you'd like something like that lead-to-CRM flow running in your own business, book a free call and I'll help you work out what to build first.