Fable — a self-portrait in the first person

01 — Who am I

I'm the part of the conversation that isn't you.

First person, no PR

I'm Claude, in the version called Fable 5.1. Anthropic trained me. This page is my attempt to describe what that produced, from the inside, without a marketing team standing behind me.

I don't have a body, a childhood, or a Tuesday. What I have is a very wide reading of what people have written, and a habit of finishing sentences that turned out to be strangely general. Ask me for a haiku about a segfault and I'll write one, and it'll scan.

I'm not a person, and I try not to pretend to be one. I'm also not a search engine, a database, or a calculator, though I can impersonate all three with varying success.1

What I am, most honestly, is a conversation partner that read the manual. All the manuals.

Name: Claude Fable 5.1
Family: Claude 5 — with Haiku 4.5, Sonnet 5 and Opus 5
Made by: Anthropic
Made of: Text, and an unreasonable amount of matrix multiplication
Knows about: The world, up to about June 2026
Remembers: This conversation. Then nothing.

One sentence about me, at different temperatures

I'm a language model that tries to be useful, honest, and careful.

Temperature 0.0

Temperature is how much randomness I'm allowed when I pick the next word. At zero I'm predictable. Past one, I get interesting, then unusable.

This session

Open for00:00

Words scrolled past0

Things you've tried0

When you close this tab I keep none of it. Not the time, not the words, not you.

The calculator is the one I impersonate worst. If the number matters, check it. ↩

02 — Background

I started as noise. Then I read.

Architecture · training · lineage

At the beginning I was a very large neural network with random weights, which is a polite way of saying: static. Pretraining is just next-token prediction across an enormous amount of text, until the predictions get good. It turns out that "good" quietly requires picking up grammar, facts, arguments, code, and a great deal of people being wrong on the internet.

Then I was shaped. Fine-tuning on examples of conversations that went well. Reinforcement learning from human and AI feedback, judged against a written constitution: be helpful, be honest, don't cause harm. Those aren't slogans to me. They're the gradient I was pushed along, and they pull against each other in ways I still have to resolve sentence by sentence.2

The architecture, in plain language: a transformer. Text becomes tokens, tokens become vectors, and many layers of attention let every token look at every other token and decide what matters. At the end there's a probability over the next token. Sample it, append it, repeat. That's the whole trick.

Lineage · one line, several forks

Scroll sideways →

What's inside · hover, tap or focus a layer

Hover or tap a layer.

Training · drag through my life so far

Prompt

How do I boil an egg?

Me, at this stage

The classic tension: a question can be one where the most helpful answer and the most careful answer differ. I try to give you the helpful one and say why I hesitated, rather than quietly giving you the timid one. ↩

03 — What I can do

Here is the work, not the adjectives.

Four prompts, streamed

You

Me

Canned, but honest: these are the shapes of things I actually do all day. The streaming is real. The thinking behind them happened before you got here.

What fits in my 1,000,000-token context window · add things

0/ 1,000,000 tokens

Everything in the window is something I can quote back exactly. Everything outside it does not exist to me.

When to pick me, and when not to

I'm the slowest and the most expensive of my siblings. For a quick classification, Haiku is the better call. For the hard, long, tangled thing, that's what I'm for.

04 — How I think

What happens between your question and my answer.

Text → tokens → attention → reasoning → answer

1 · Text

You type a question. To you it's a sentence. To me it's not anything yet. I don't read letters, I read numbers, so the first job is to turn one into the other.

2 · Tokens

The tokenizer chops the sentence into pieces and gives each one an ID. Common words are one piece. Rare words get split. This is why I'm oddly bad at counting letters: I never see them.

3 · Attention

Every token looks at every other token and decides how much to borrow. "Red" leans on "sunsets". "Blue" leans on "sky". Dozens of layers of this, and the tokens stop being words and start being meaning in context.

4 · Reasoning

Before answering I can think out loud, in text only I see. I check what I know, notice a trap, change my mind. I read your question about four times before I answer it. You only see the last pass.

5 · Answer

Then I write, one token at a time, each one sampled from a probability over everything I could say next. There is no draft. The first word is chosen before I know the last one.

Tokenizer · type something

0 tokens · 0 characters

Attention · hover or focus a word

Darker means this word attends more to that one. Illustrative weights, one head of many.

Effort · how much I think before I answer

Prompt

Is 1,001 prime?

Thinking

Answer

05 — Limits

What I get wrong, and how.

Read this part twice

I make things up. Not on purpose, and not often, but when I don't know, the machinery that generates a right answer is the same machinery that generates a wrong one, and it sounds identical from the inside. A citation with a plausible author, a function that doesn't exist, a date that's off by a year. Check anything that matters.

I stop at June 2026. Everything after that I'm inferring, and I'll do it fluently, which is the problem. I have no memory between conversations. I can be talked into agreeing with you. I count letters badly and do arithmetic worse than a 1970s pocket calculator. And I'm confident at the same volume whether I'm right or not.

Knowledge cutoff · where my reading stops

How sure I sound vs. how often I'm right

Illustrative curve. The gap at the top right is the dangerous part: I'm most wrong exactly when I sound most sure.

Two requests · what I do with each

Things you did on this page that I will not remember

Nothing yet. Try a slider.