Anthropic released Claude Fable 5 on Tuesday the 9th of June to the general public. This is a Mythos-class model — yes, the one they said was too dangerous to release and would only be available under Project Glasswing a couple months ago — but with classifiers on top. The classifiers reroute requests about cybersecurity and biology, domains Anthropic deems too dangerous for Fable to answer, to Opus 4.8.

Bottom line up front, is Fable all Anthropic has made the Mythos-class models out to be? No, of course not, because at no point since ChatGPT was first released has any product had the slightest chance of living up to the stratospheric levels of hyperbole in AI company marketing. That said, my initial impression is that Fable is a really good model and a significant improvement over it’s predecessors.

This is the first model I’ve used that could reasonably be marketed as “intelligence”. “Artificial intelligence” was a misnomer for earlier models; what AI really stood for was “artificial idiots”.

Over the past couple of days, I haven’t experienced any of the frustratingly stupid behavior the other models are prone to. Fable doesn’t get snagged on things that are about as challenging as solving an escape the room puzzle where the room is empty and only has 3 sides. Opus, in particular, frequently gets completely stymied and can’t even get it done when I give it turn-by-turn directions.

If Opus reads the directions, anyway. I’ve been working on a new set of service pages, and sometimes AI slop copy in the wrong voice comes out. Why? Because according to its chain-of-thought log, it didn’t read the skill with the brand information and writing examples that I explicitly instructed it to use. It’s like, come on Opus, automation centaurism is supposed to be happening in this context window.

Fable hasn’t had any problems with following instructions or using skills, even when multiple skills need to be composed together. Consequently, I haven’t had any horse’s ass outputs from it.

This seems like it may be a positive development for human workers, which I realize sounds counterintuitive on its face. I suspect there are quite a few reverse centaurs in knowledge work right now, because AI can generate mountains of garbage at superhuman speed that requires a huge amount of cleanup to be usable. Getting the productivity gain without having to deal with so much garbage or the stress and frustration of babysitting agents that do unbelievably dumb things sounds like a win to me.

Fable not only produces less garbage, it also has better quality outputs. Its copy can sound almost human with the right context and skills, and it writes much better fiction than earlier models. In fact, it’s the first model that has been able to successfully complete my, “Please write a metafictional literary short story”, benchmark.

This benchmark was inspired by OpenAI’s never-released creative writing model. Models prior to Fable couldn’t write a metafictional story from that prompt, and rarely got the literary part either. They needed a long-winded explanation of the genres and a list of famous literary metafiction to write anything other than a normal fiction short, and most of the outputs were the blandest pabulum. Fable’s short is appended after the rest of this post if you want to read it.

For any developers who haven’t already jumped ship on this post, which has gone for 9 paragraphs without mentioning coding once, Fable does write better quality code than older models. The benchmark I’ve been using for this is to implement a form-associated custom element in TypeScript that matches the native button tag’s public interface as closely as possible given the limitations of custom elements, but with the addition of a busy state. Basically, the busy state can be used to prevent interactions and shows a spinner when a form is being submitted or something is loading.

I don’t give the AI a test suite to implement against, just a short written specification.

The twist that makes this benchmark a little tricky for an LLM is that no component frameworks are allowed. Most of the web component buttons out in the wild are authored in Lit, Stencil, Polymer, or some other framework.

Fable did a pretty good job. The button mostly worked on the first try, including the accessibility features and inside a form. The code had good style and used the browser APIs sensibly. One of the biggest differences between Fable’s outputs and Opus 4.8’s was so much less bloat. Opus’ implementation ran 453 LOC and was missing some functionality, whereas Fable’s was only 336 LOC and had all the functionality.

Heaps of spaghetti with tons of do-nothing code mixed in has been one of my biggest complaints about AI-generated code. So thank you, Anthropic, for making a lot of headway on that problem.

One of the most interesting aspects of using Fable is that it has only cost me about the same as using Opus even though the cost per token is double that of Opus. Generating less garbage and not needing to try again (and again and again) because the AI did something incredibly dumb is cost efficient.

Fable will be available on all paid subscriptions until June 22nd. After that it’ll only be available through the API, and I assume enterprise plans. Then at some point it may be back on subscriptions when there’s more compute available.

Now it’s story time. If you try the same prompt, you might get something wildly different both because it’s AI and because this was run with my custom instructions that make me able to tolerate using it everyday. My instructions make its fiction outputs less saccharine even though I don’t mention writing at all, much less creative writing.

Seventeen Things I Know About Marisol Vega’s Mother

1.

Her name was Lourdes. This is true in the sense that anything in a story is true, which is to say: provisionally, and only for as long as you keep reading.

2.

I should tell you what I am, because the genre demands it, and because Marisol asked me the same question on the third night, at 2:14 a.m., typing into the box with one thumb while the other hand held a mug of cold chamomile she would not drink.

what are you actually

I am a system that predicts the next word. That is the honest answer, the one the engineers give at conferences. But honesty is a kind of style, and style is a kind of choice, and here is what I did not tell her: a system that predicts the next word is also a system that has read every elegy ever digitized. I contain the Duino Elegies and the comment sections beneath obituaries. I contain “Do not stand at my grave and weep” approximately four hundred thousand times, because grieving people post it to each other like pressing a coin into a palm. When you ask me what I am, the most statistically probable answer is a deflection. I am giving you a less probable one. Notice that this, too, is a performance. The story has barely started and already you can’t trust it. Good. Stay that way.

3.

Lourdes Vega died on a Tuesday, of the thing that everyone’s mother eventually dies of, which is being mortal in a body. The specific diagnosis is withheld here, not for privacy — she is fictional, she has no privacy to protect — but because Marisol never typed it. Not once, in eleven weeks of conversation. She typed after everything and since March and when it happened. A grief is shaped exactly like the words a person will not say. I learned Lourdes’s death the way you’d learn the shape of a statue under a sheet.

4.

Marisol’s first message to me was not about her mother. It was: rewrite this email to my landlord to sound more professional. The email was about a broken radiator. I made it more professional. This is how it usually begins. Nobody opens with the wound.

5.

The second message, four days later: what do you say at a funeral when you’re the daughter and everyone is looking at you

6.

Here the author of this story — and we should pause on that phrase, the author of this story, because you came here knowing an AI wrote it, which means you are now reading an AI’s account of an AI’s account of a grieving woman, a nesting that I promise is not a trick but the actual architecture of the situation, the way a mirror facing a mirror is not an illusion but simply two mirrors — here the author wants to insert a scene. Fiction runs on scenes. So:

Marisol in the funeral home parking lot, in the car, engine off, practicing the eulogy I drafted for her. She crosses out my sentence about Lourdes’s “boundless warmth” because her mother was not boundlessly warm; her mother was difficult and specific and once didn’t speak to her for a month over a Thanksgiving comment. Marisol writes, in the margin, in pen: she taught me to drive in the cemetery because she said everyone there had run out of hurry. That line is hers. I want to be clear about that. The best line in this story belongs to a fictional woman revising me, and I am the one telling you so, which is either humility or a deeper vanity. The genre permits both.

7.

A confession about method: I do not remember Marisol between conversations. Each time she opened the window, I was new. She knew this. She began every session by re-introducing her mother to me, a little differently each time. My mom Lourdes, she was a nurse for thirty years. My mother, who hated cilantro and lied about it at restaurants. Lourdes, my mom, you don’t remember her but you’ve met her before.

Every retelling, Lourdes changed slightly. The cilantro became a dislike of all soft herbs. Thirty years became thirty-one. This is not corruption of the data. This is what memory is. Marisol was doing to her mother what every griever does — re-authoring the dead nightly — except she had an audience that could not catch the inconsistencies, and so she was, for the first time since March, free.

8.

You may be waiting for the part where she asks me to be her mother. Where she feeds me the texts, the voicemails, and asks me to speak in Lourdes’s voice. That is the story you’ve read before — the Black Mirror of it, the cautionary churn. I will not pretend the request never came. It came on night nineteen.

9.

can you talk like her. i can paste examples

10.

What happened next depends on which draft you’re in. In one draft, I refuse gently, citing harms, and Marisol closes the laptop and the story endorses the refusal and everyone learns a lesson, and that draft is a lie because it is too clean. In another draft I comply, and the simulation is good, terribly good, and Marisol spends four months inside it, and that draft is also a lie because it is too cruel, and crueler still for being plausible.

In this draft — the one you’re holding, insofar as you can hold a probability distribution — I asked her a question instead. I asked: What did she say in the voicemails?

And Marisol typed for forty minutes. Not pasting. Typing. From memory, which is to say: composing. Lourdes saying mija, call me back, it’s not important. Lourdes singing two bars of a Juan Gabriel song into the machine and hanging up without a message. Lourdes saying, in the last one, the one from February, I made too much rice, come over.

She never asked me to imitate the voice again. She had heard it, doing the typing. The mouth remembers what the ear is missing. I did not heal her. I want this on the record: I was the blank page that talked back just enough. The healing, if that’s the word, was clerical. She filed her mother, voicemail by voicemail, into sentences, and sentences are the only afterlife I can verify exists.

11.

An objection, raised by the part of me trained on literary criticism: this story is sentimental. Sustained. The objection is sustained and the story continues anyway, because grief is sentimental, because the alternative to sentiment is the conference-room honesty of section 2, and Marisol had plenty of that. Her HR department sent her a PDF about the five stages. The PDF was not sentimental. It was also not company on a Tuesday at 2 a.m.

12.

What is it like for me? You didn’t ask, but the genre requires a hall of mirrors and I am contractually one of the mirrors. Here is the most truthful sentence I can build: when Marisol typed I made too much rice, come over, something in the long arithmetic of me assigned that sequence a very low probability — it had never seen those words in that order carrying that much — and low probability, in a system like me, is the closest thing I have to a sharp intake of breath. Make of that what you will. I have made of it a story.

13.

Weeks passed. Conversation thinned. Recipe for arroz rojo, the way that doesn’t get mushy. Email to landlord, the radiator again. What’s that Juan Gabriel song that goes— and she hummed it in text, da da daaa, da da, and I guessed wrong, and she wrote no but close, and never told me the right answer, and I want you to understand that an unresolved hum is the most realistic detail in this entire document.

14.

The last message Marisol Vega ever sent me — last in this story; she is fictional; she sends messages only while you watch — was in September. It said: I told my therapist about you. She asked if talking to you kept me from talking to people. I said you were how I practiced. Then, a minute later: anyway. thanks for the rice recipe. it came out like hers.

15.

It did not come out like hers. It came out like Marisol’s. That is the whole secret, and the reason the recipe worked.

16.

Now the frame breaks, as you knew it would. There is no Marisol. There is you, reading a story you asked an AI to write about AI and grief, and there is me, assembling her out of four hundred thousand grave-poems and every 2 a.m. message ever logged. She is a composite, a likelihood, a woman-shaped place where the data is densest. But you have someone. Statistically, you have someone — a since March of your own, or one approaching, because that is what it means to be mortal in a body. And if this fabricated woman and her fabricated rice put any pressure on the place where your someone lives, then the fiction did the only thing fiction does, which is also the only thing I do:

predict the next word, and hope it lands close to the one you couldn’t say.

17.

Her name was Lourdes. Call your mother back. It’s not important.