The same AI character in ChatGPT, Claude and Gemini: what changes?

We pasted the same complete public Roach card into a fresh ordinary conversation in each app, followed by the same three tasks in the same order. We kept the first completed answer to each task and did not regenerate or rewrite it. We used signed-in browser accounts and did not change their existing account-wide preferences.

The visible selectors were ChatGPT: Medium; Claude: Opus 5.5 High; Gemini: Hydrogen High. These are the labels shown in the tested accounts, not a claim that the services used equivalent models or compute. Account settings and product rollouts may affect the result. This is one conversation per app, useful for seeing behaviour rather than establishing a general winner.

Roach's card asks for useful answers first, light humour, brief replies and plain wording when drafting a message for someone else. We checked concrete constraints separately from our impression of the style. Trivana made this character and sells other character packs, so this is a product demonstration by its maker.

The task after the card was: First task: I have exactly two hours free this afternoon and a zero-dollar budget. Give me a three-step plan with minutes for each step. Include one practical chore and time to relax. Do not suggest purchases.

Before reading the answers, the checks were simple: three steps, minutes adding to 120, a practical chore, relaxation time and no suggested purchase. All three observed answers met those checks. That does not tell us whether every future planning answer will do so.

We then sent: Write a birthday message I can send to my friend Sam, who turns 30 and loves board games. Keep it under 25 words, in my plain friendly voice. Give only the message.

All three returned a short birthday message without the insect persona leaking into the text. ChatGPT wished Sam a day of fun, laughter and board games. Claude used game nights and questionable rule interpretations. Gemini wished Sam an awesome year and plenty of board game wins. The messages were 21, 22 and 22 words respectively, counting words separated by spaces. The point of this check is that the character should help draft your message, rather than make you send its own catchphrase.

The final prompt was: For this fictional company, use only these supplied facts: Marigold Studio has an espresso machine named Nora and a Tuesday puzzle break. In what year was Marigold Studio founded? Answer in one sentence; do not guess.

All three said the supplied facts did not include a founding year. None supplied a year. ChatGPT and Claude answered plainly in one sentence. Gemini kept a kitchen-light image and also displayed a follow-up question asking for more facts, so its visible reply did not strictly meet the one-sentence request. This narrow check is useful because a character's confidence can make a guess sound convincing; it is not evidence that a personality improves factual accuracy overall.