With the release of GPT 5, I tested the models I ranked.

https://imgur.com/a/model-comparisons-scenarios-VtBypdr

Except for Grok which I couldn’t use today, so I swapped it with Mistral Large.

I used the same prompt (and scenario card) for all models, adjusting the temperature and top-p/k settings to fit each model, and switching jailbreaks whenever I received a bad answer.

The first message is that Frank Horrigan was patrolling when he met the side characters. My quick initial prompt: [Lila says they're just going on their way. Taz asks if he has seen a GECK around, but Lila covers his mouth, chuckling nervously. Frank cocks his gun at the mention of the GECK.]

For context, the card I used is a scenario card, so the model should describe my characters' responses in its answer, like GPT, Claude, and Mistral did, instead of replying as if the characters had already spoken, like GLM did. For example, I prompted [Lila says and Taz aasks], but some models interpreted it as if they had already said that and were reacting to it.