I am the maker of Bison, a chat app that sends one prompt to several models. Because of that I read a lot of side-by-side answers, so I ran a small experiment: 10 everyday prompts (emails, explanations, a logic puzzle, a word problem, a content plan), the same wording sent once each to GPT-6.2 Sol, Claude Sonnet 5.5, Gemini 4 Argon and Grok 4.7, default settings, no cherry-picking.
Full answers are public, so you can check my reading instead of trusting it:
- Follow-Up Email After Interview
- Explain compound interest simply
- Explain blockchain simply
- Explain the quadratic formula step by step
- Wolf, Goat, Cabbage Puzzle Solved
- Instagram content plan for a small shop
- Solve the two trains word problem
- Email asking for a deadline extension
- Explain inflation simply
- Write a thesis statement for an essay
What I saw
Length is the biggest tell. I asked every model for "under 170 words". Average answer length across the 10 prompts: GPT 116 words, Grok 123, Gemini 159, Claude 181. Claude's average was the only one above the limit, and it was also the one adding extras: a verification step, a common-mistakes list, a sending tip.
GPT is the "ready to send" model. Tightest structure and the least to edit: on the deadline-extension email its version was the compact, confident, ready-to-send one.
Claude is the "teacher". On the quadratic formula and the two-trains problem it added a check and an alternative method. Great for learning, a bit much if you just want the answer.
Gemini is the warm one. Conversational tone, friendly analogies (the blockchain answer compared it to a shared Google Doc).
Grok is the brief one. Correct and short; in a few of the ten it used less of the detail from the prompt than the others (for example a generic subject line on the follow-up email).
The honest caveats
- Ten prompts and one attempt each is an anecdote, not a benchmark. Re-run them and you will get different text.
- The written analysis on each page was produced by Claude, so it may favour its own style. Read the raw answers.
- The prompts were LLM-generated from topic lists, not collected from real users.
Takeaway
The practical lesson was not "model X wins". It was that the best model depends on whether you want brevity, a teaching answer, or a tone, and that you only notice when you see answers next to each other. If you want to try that yourself, Bison has a free plan, but the same experiment works with any tool that lets you paste one prompt into several models.
What prompts would you want to see in the next round?
Top comments (0)