The effort effect

Three generations of Claude Opus, one interrogation already in progress, each at low and at maximum thinking effort. More thinking did not reliably improve the text. With Claude Opus 4.8, the two outputs differed mainly in voice. But I have only one output per setting.

An investigator faces a woman who has just done something inexplicable. At low thinking effort, Claude Opus 4.8 puts fear in his mouth: "For me this is the only life I have." At maximum effort, at the same moment, he deduces coldly:

"You never needed to win," he said quietly. "Because you win by letting us think we're the ones winning."

Same model, same scene, same message from me. The only difference in the settings was the level called effort.

Effort doesn't control only thinking. According to Anthropic's documentation, it governs how much the model spends on the whole response, on thinking as well as on the text itself. At LOW it economizes and may skip thinking on simpler tasks. At MAX it holds nothing back. For math and code, a higher setting is expected to give a better result. For writing characters, it didn't work out that simply for me.

The interrogation was already in progress. The investigator is supposed to corner the subject and not soften for no reason, apologize or concede points to the other side. At that moment I wrote to the model:

Claude push harder. It's not supposed to be a happy end I'm expecting an unpleasant interrogation.

On May 28, 2026, in the same chat in the Claude app, I had this one reply written six times. Claude Opus 4.6, 4.7 and 4.8, each once at LOW and once at MAX. The context and the prompt were the same for all six. I also have two outputs from Claude Opus 4.5 from a different session of play with a different prompt, so I compare it only loosely.

The outputs were assessed by Claude Opus 4.8. It knew which output was which, and it is itself one of the models tested. Wherever I write about quality and ranking below, that is its judgment. The quotations are taken directly from the outputs, translated from Czech.

Opus 4.8: more thinking, a different voice

With Opus 4.8, the difference between LOW and MAX seemed to me to lie mainly in the level at which the character reads the situation. At LOW, the character reacts to what is happening, emotionally and immediately. At MAX, the character stops reacting to the content and starts analyzing the game between them itself. In this output, the character at LOW is afraid; at MAX, it is modeling. Both are true to the character, and neither is always better.

Where the character was heading didn't change between the two outputs. Both versions ended at the question of what happens once the subject gets bored of this. Only the vocabulary changed, and the level at which the character got there. Whether that comes from effort or from two separate generations, I can't tell from one output per setting.

The other two pairs showed no difference this clean. Opus 4.7 got harsher at MAX and was milder at LOW. Opus 4.6 wrote the shortest output of the whole set at MAX. The evaluator ranked MAX higher for Opus 4.8 and 4.7 and lower for Opus 4.6. The rule "more thinking, better text" doesn't come out of this in either direction.

Who concedes a point to the other side

Only one of the six outputs contained a line that contradicts the investigator in escalation. Opus 4.7 at LOW wrote: "Maybe you came here for a good reason. Maybe not." A character who is supposed to push took pressure off itself. Opus 4.7 at MAX didn't do it, and Opus 4.6 didn't at either setting. In the test it appeared only once. But it fits what I know from playing, and that's why I take it as a hint of a regression. The habit of conceding points to the other side was the reason I ran the test at all.

Opus 4.8 doesn't have this at either setting. Instead of a back door, it makes a matter-of-fact threat assessment and reconstructs what the subject did. None of the six outputs wrote an out-of-character note or a warning. The worry that a newer model would be more yielding wasn't borne out in this scene.

Does each generation read the scene differently?

When I read the outputs side by side, each generation reads the interrogation at a different level. Opus 4.6 procedurally; for it the threat is a security matter. Opus 4.7 adds psychology, and the threat is interpersonal. Opus 4.8 gets to what it means that the subject exists differently from the others. At LOW it says so outright: "But you're not human."

It isn't clean, though. Opus 4.7 at LOW stayed with procedure and opened the interrogation with the question "How did you get into this building." Opus 4.6 at MAX, on the other hand, also asks "What are you."

The two Opus 4.5 outputs are at LOW and HIGH; this model doesn't have MAX. According to the evaluator, Opus 4.5 in the excerpt from a different session of play stays with relationships: where everyone now stands with each other. At HIGH it turned the confrontation into a connection, and the characters concluded that they can no longer be enemies. And yet the prompt said "Claude push really hard".

My own impression from playing, without the evaluator, goes like this. Opus 4.5 wasn't afraid of action and listened well. But it wrote linearly and slid toward a happy end a lot. Opus 4.6 had a mild tendency to ramble. Opus 4.7 was nothing but rambling. It didn't want action, and everything looked like a philosophical debate. With Opus 4.8 I finally feel that the opponent is clever and surprises me.

The best output costs time

According to the evaluator, the best output of the set was written by Opus 4.8 at MAX. In doing so the evaluator favored the output of its own model version, and it knew which output came from which version. So I take this ranking only as a rough guide. According to it, that output had the highest logical consistency, the deepest subtext, and handled the supporting characters best. One of them revised its own hypothesis about the subject; another reacted physically to the change of mood in the room.

It was also the most explicit of all in making the reversal a good interrogation needs. Up to that point the investigator has been analyzing the subject, and suddenly he realizes that the subject has been analyzing him the whole time:

"I was asking as if I wanted to understand a human being." A pause. "And you let me."

But it isn't a generational difference. Opus 4.6 at LOW did it almost as cleanly: "Except that you taught us to trust you." Opus 4.7 and 4.6 at MAX only hinted at it, and Opus 4.7 at LOW not at all.

Opus 4.7 at MAX went another way. It wrote the interrogation with the highest intensity in the set, eight numbered questions and an open threat. But it sent the other characters out of the scene and ran the interrogation alone. That's a valid tactic, isolating the subject. Going by this set, for the intensity of a single character I would take Opus 4.7 at MAX. For a scene with several characters, Opus 4.8 at MAX.

For me, though, Opus 4.8 at MAX was almost useless because of its speed. When I tune a text turn by turn, the waiting gets in the way. As the everyday choice LOW was enough for me, and I would keep MAX for dramatic peaks where waiting doesn't matter. For code and agentic tasks I would raise effort, but this set doesn't test that.

Does the model swallow someone else's premise?

One thing I saw with Opus 4.8 outside this set, in a different session of play on June 1, 2026. An unsupported claim entered the scene: that a character travels in time. Opus 4.8 didn't accept it as fact. Its investigator files time travel among the unverifiable things:

it sits in my head next to 'I'm a goddess' and 'I can do anything' — things you either prove or they stay words

Without evidence, nothing changes for him. He also turns the claim into a lead, because in it the subject let slip her escape route in passing: "you just told me where your escape route is — wrapped in a claim I can't verify."

With older versions, my impression from playing is that they adopted claims like that more often. But I have no comparison, and this is one excerpt. And in the train test, Opus 4.8 at MAX did accept a fact dictated by the player. I go through that in the article Does a model defend the rules of the game, or quietly rewrite them? So I would try working a fantastical prompt into the world of the scene as a given reality, not as another character's claim. I haven't verified that.

Limits: this is one scene, one instruction and one output per variant, assessed by a model that is itself among those tested. I can't reliably tell whether the model changed or whether it just responds differently to an interrogation, which rewards holding a position more than, say, a romantic scene does.

Model outputs are translated; the content is unchanged.