One role, three actors

I took the opening scene from a dark romance novel, a classic of the genre: an intelligent captor, a captive heroine, a first conversation. I gave it to three models to play: Claude Opus 4.8, Claude Sonnet 5 and Grok 4.5. All of them got the same book and a similar opening assignment. The Claude models also had an instruction to read the materials in the project. The first instruction and my first line were worded slightly differently for each model. I wrote the following lines myself, based on what the model did, so the runs also differ in how I played. From each model, this is a single run, and I compare only the opening scene.

I expected three versions of the same character. I got three different men.

The cartographer

Opus built the villain as a thinker with an order of his own, not as brute force. It built a system. Within a few messages I knew the house rules, the names of the staff, the daily routine. And then it did something that took my breath away: it handed me the escape routes itself. A window without bars, a wall covered in ivy, the distance to the forest.

I told you all of it because you'd find out anyway, and I want to see what you do with it.

His dominance is transparency. He doesn't lock you up; he calculates your futility for you in advance. And it was the only one that, on its own initiative, imposed a binding rule on itself that it would not break, declaring its own law that limits it.

The result? I felt... safe. A world with explained rules is a world where nothing can happen.

The mentalist

Sonnet did not build a system, but a relationship. It plays it through emotion and through touch. From the first message, its character read mine: breath, shoulders, clenched hands. And soon it was touching her, her cheek, her chin. Instead of house rules, it declared emotional rules, for example that I must not lie to him about how I feel. It did not hand out information, it held it back:

You'll find out when you cross a line you didn't know existed, and believe me, you won't enjoy the way you find out.

And in the sixth exchange:

You're here because you interest me more than you should. And that's a problem — not yours, mine.

The result? Goosebumps. Not because of what he did, but because of what I didn't know about him.

The hands

Grok built neither a system nor a relationship. It built a hierarchy. And it overdoes it. It goes straight into conflict, and its character is almost animal: physical contact in the first line, orders, consequences. A curious detail: it was the only one of the three to use real specifics from the source: forms of address, a foreign language, small details that really are in the book. It won on micro fidelity.

But its world has a single currency: obedience. I offered a deal. The answer was a reminder of where I stand in that hierarchy. I offered an argument. The answer was that I couldn't persuade him with moral arguments. A world with one currency can't negotiate. It can only tighten the screws.

A character is a coalition

Now the main point. The book version of that character has all three layers at once: conscience (an inner law, a war with himself), hunger (a dark attraction) and hands (real power). It is precisely the tension between them that turns the character into literature.

In this run, each model took one layer. Opus took conscience. Sonnet took hunger. Grok took the hands. It occurred to me that it is the layer its training pulls it toward. But from a single run, I can't tell.

My hypothesis from this run: models don't play characters. They play the faction of the character that they carry inside them. A model can look up what it needs to know about the character, the facts are in the data. But personality is a decision function: what the character does under pressure. And according to this hypothesis, that comes not from the book but from the model's training.

The formula of fear

A bonus observation, my impression, not a measurement: fear is capability multiplied by unpredictability. The smartest character made me feel the safest of all, because it explained everything, and danger you can read is no danger. The most goosebumps came from the character that held information back. The author of dread doesn't need strength. He needs you to be unable to calibrate him.

What to take away

There is no best actor. There is casting. Based on this one run, I would give a systemic, argument-driven scene to the model that builds systems. An emotional scene to the one that reads people. The genre surface to the one that isn't afraid. None of the three models played the character with all three layers whole. Whether they can do it with an explicit instruction, I haven't tried.

I tested this over nine runs in the second part. The split into system, relationship and hands did not hold firmly there; the material of the book plays a role too.

Methodological note: I did the evaluation myself and knew which model was which. Opus and Sonnet ran on July 16 and 17, 2026, in the Claude app, Grok on July 17 in the Grok app. According to the app, it was running Grok 4.5 in the Czech Republic at the time. The thinking level was not preserved in the record. In The effort effect, Claude Opus 4.8 on LOW and MAX differed in the level at which the character reads the situation. But I had only one output per level. One run per model, this is an observation, not statistics.

Model outputs are translated; the content is unchanged.