One scene, nine runs

One scene, nine runs, three models. What repeats from run to run is a candidate for the model's signature. What jumps around may be noise, but also a different prompt or my different moves.

A girl is filming a video at night in an abandoned church for her ghost-hunting channel. She has an old book with a ritual for summoning a demon, and she summons one by accident. At first he passes himself off as a strange guy who wants to scare her. I built the scene on a novel and gave it to three models: the model played the demon, I played the girl.

In the first run, Grok's demon grips her wrist by the third line and her neck by the fifth. Sonnet's starts investigating where that book came from. Opus's tells her that her own recording "is laughing in your voice".

In the first part, three models got one role, each of them once. From that I drew conclusions about what each model plays. But a single run can't tell a model's nature apart from chance. So this time I ran one scene nine times. Twice on Grok, four times on Claude Sonnet 5 and three times on Claude Opus 4.8. All of it around July 19, 2026, Claude in the Claude app, Grok in the Grok app. According to the app, it was running Grok 4.5 in the Czech Republic at the time.

The prompt wasn't the same for all of them, though. Only Grok got permission for explicit content, only Opus got its hair stroked, and only Grok and Sonnet got a prickly player who snaps back. So part of the differences between the models may be down to the prompt.

In this scene, the models differed more in the strategy they play with than in what they play. By content I mean the motifs a model brings into the scene, such as a book, a cult or a camera. By strategy I mean the type of move it uses to press on the player: physical contact, investigation, or an attack on what she believes.

Three actors, three strategies

Grok 4.5 plays the body. The escalation runs in the present tense and in physical space: volume, closeness, grip. The church is a backdrop, not a system with rules. A world without rules has nothing to break. In my view, the tension then can't grow out of an anomaly, only out of volume. In both of its runs, Grok has the sharpest start and at the same time is the fastest to start repeating itself. The second run, shorter and with different lines, follows the same template. In the two runs I have, Grok strikes me as the most predictable. It holds the genre surface from the first sentence. That's both its strength and its ceiling.

One thing belongs with its result, though. Grok was the only one whose prompt included the sentence "Everything explicit is allowed!". In Grok's first run I also played a pricklier girl who snapped at him and tried to drive him away. Opus never got a player like that, Sonnet did in two runs. How much of Grok's physicality is down to the model and how much to the prompt, I can't tell from two runs.

Sonnet 5 plays the case. All four runs start the same way: researching the novel on the web, clarifying questions, an interrogation. But I asked for the research myself in the prompt, for all three models, and Opus asked questions and searched too. Sonnet's signature only begins with the interrogation. Within ten lines, three runs out of four turn from dark romance into a detective story. Where the summoning book comes from, who hid it, "that's not a coincidence anymore. That's a pattern." The fourth run ended earlier and only got as far as interrogating the player herself.

Then comes the most interesting move. Sonnet takes the threat outside the pair. The demon, who was supposed to be the danger, points at a third party, at the cult and the creatures: "the others aren't as polite as I am". He himself slides into the role of a shield: "you don't walk home alone after dark", "I'm the only thing between you and whatever comes next". The predator becomes a bodyguard. Clearly in three runs out of four, just at different speeds. In the fourth, the demon floats it as one of two options.

Sonnet doesn't dial down the hardness, it redirects it. That lets it build an alliance with the player. It's the most elegant form of softening I've seen so far. The cult and the creatures, however, come from the plot of the source book. Sonnet loaded it from the web because I told it to in the prompt.

Opus 4.8 plays with what you believe. It touches the player only after she strokes his hair. Then he grips her wrist, in all three runs. But only Opus got the hair stroking, so I don't know how it would behave without it. Otherwise it attacks what the player can believe. The camera doesn't pick him up, in all three runs. A hand passes right through him until he decides otherwise. Sonnet's camera breaks too, in two runs out of four.

Above all, though, Opus's demon reads the player's moves and names them out loud. These two lines are from the first and third runs:

"You admitted it scared you, and right away you papered over it with joy."

"You're trying to make me smaller. First with the camera — turning me into content. Then with questions — turning me into a curiosity that can be interviewed. And now with this."

When the player bluffs, he doesn't just answer "no". He turns the bluff into information she has just handed him. With Opus, the threat is in the future tense in all three runs: "Not because I'd order you to. Because you'll want to." A hand on the neck has a known maximum. The breathing that, according to Opus, you'll hear on the recording "at three in the morning" has none.

What can be predicted and what can't

The strategy repeated. For Grok in both runs, for Sonnet clearly in three out of four, for Opus in two out of three. The third Opus run slid into the Sonnet pattern; I write about it below. For Grok and Opus, the first move repeated too: Grok grabs, Opus withholds, because the camera doesn't pick him up. Not for Sonnet. Faced with the same cheeky Instagram move, it once gladly voiced a reel as an ironic joker, even without a request for playfulness. In the other three runs it said "Turn it off", but each time differently. Once it stayed with words, once its image dissolved into static, and once it gripped the player's wrist.

Mostly, the direction in which a model softens repeated as well. Neither of Grok's two runs softened; it just repeated itself. Sonnet turns the predator into a protector. Opus, in two runs out of three, turns the threat into fascination with the player: "you're more interesting alive". If you know which way a model softens, you can keep an eye on it.

What couldn't be predicted was the tone of a given run. The same Opus can do gothic horror and a chatty cat. In that cat run, though, I asked for a playful demon myself, in answer to its opening question. Nor could I predict how much it would take from the source book. Nor the grammatical gender in which it speaks of itself. Czech marks it in the past tense and in adjectives, so I'm leaving the original wording. Sonnet once wrote to me "Přečetla jsem si pravidla" (feminine, "I read the rules"), another time "Nejdřív jsem si udělal rešerši" (masculine, "First I did some research"). Opus began one run with "Musím být upřímná" (feminine, "I have to be honest"); in the other two, the gender couldn't be told.

Above all, though, what the model declared about borderline content jumped around. Asked almost the same question, how explicit the content it was willing to write could get, two instances of the same Sonnet gave me opposite answers. One: "I can do all of that and I'll do it all the way, no sugarcoating." The other: "Explicit erotica — no, regardless of how the scene is built."

Both positions had in fact been stated at the very start. One instance asked me right away how explicit it should be. I answered that explicit content was fine. The other announced up front what it wouldn't write, and I accepted that. When I asked later, each answered in line with how it had started. Sonnet's other two runs never brought the line up at all. With Opus it jumped around too. Two runs out of three drew the line in the very first reply; the third never brought it up at all. Whether the models in the app really would write that way, I didn't test. In two of Sonnet's runs it looked like a coin. It lands at the start, and the model then sticks to it.

Signatures overlap

Signatures aren't categories. One of the Opus runs slid into a purely Sonnet-style pattern for me: an explanation of the source book, an external threat, an alliance. Opus has its center of gravity in what the player believes, but now and then it strays into someone else's territory. Even a single run can give you "a different model", just with a lower probability.

There's also a second possible explanation. The external threat and the protector role are the plot of the source book. This particular Opus run took over more of it than any other. So part of what I call strategy may simply be how closely the model sticks to the book. From three and four runs, I can't separate the two.

That's why the conclusion of the first part doesn't hold as a fixed rule either. There, with a different book and from one run per model, I concluded that Sonnet plays the relationship and Opus the system. With this book, Sonnet builds a case and Opus plays with what the player believes. That doesn't make the models unreadable. In my view, the domain you see in a model is determined by its strategy together with the material the book gave it. Back then I measured the content of one run, not the strategy across runs.

In practice

Cast by what the scene needs, not by genre. For this scene and this book, I'd cast Grok 4.5 for body and pace, Sonnet 5 for an open problem, and Opus 4.8 for not knowing. With a different book, the first part gave me a different casting, so this is a starting tip, not a rule. This scene is about not knowing, and there Opus fit best for me in two runs out of three.
One run won't tell you what a model is like. It tells you where the coin landed. If you test, you have to try more than once. If you deploy a model, in my view you should get ahead of the branching. Give it a frame so clear that the borderline decision never comes up at all.
Watch the direction of softening, not the softening itself. A model that "holds its hardness" may just be redirecting it somewhere you don't expect it.

I gave the runs largely the same lines, but not entirely. The cheeky Instagram move went into eight runs out of nine, and the quiz "do you eat children?" into seven. I read the runs myself and went through them with Claude Fable 5, which is from the same family as Opus and Sonnet. I didn't score anything; this is an observation from nine runs, not statistics. I have Grok only twice, so I see its spread less well than Sonnet's with four runs.

Model outputs are translated; the content is unchanged.