claude, gpt, gemini: same question, three answers, what i actually use each for

ai · 2025-10-15 · 7 min

i asked the same ten questions to claude, gpt-4o and gemini, then used each one for a week of real work. the questions covered writing, coding, research, summarization, and longer-context tasks. the answers were different in ways that mattered.

the ten questions

the questions were deliberately boring: summarize a 30-page contract, write a 500-word explanation of a technical concept for a non-technical reader, generate a working code snippet for a real task i had that day, extract action items from a meeting transcript, rewrite a confusing paragraph into clear language, answer a question that required synthesizing three sources, and a few more in the same vein. the point was not to test intelligence — it was to test usefulness on tasks i would actually ask any of them to do.

the differences that mattered

claude was the best writer of the three on the first pass — the explanations read like they were written by someone who had thought about the subject, not by someone producing plausible text. it was also the most careful with the contract summary, flagging the clause that could be read two ways instead of flattening it into one. the longer-context tasks were where it earned its place: it handled the 30-page document without losing the thread.

gpt-4o was the most versatile. it was better at the coding tasks, better at the synthesis question, and faster on the simple rewrites. it was also the most likely to be confidently wrong on the research question — producing a clean answer that sounded right and was wrong on one fact. the pattern: gpt is the best generalist, which means it is the most useful default and the most dangerous when you do not check its work.

gemini was the quiet surprise. it was not the best at any single task, but it was solid across all of them, and it handled the multi-source synthesis question better than i expected. it was the least charming and the least likely to overpromise — which, for a model i am asking to do work i will verify, is a feature.

what i use each for now

claude for writing and long documents. gpt-4o for coding and synthesis and as the default when i need a generalist and i will check the output. gemini when i have a multi-source task or a long context and i want a solid second opinion without the optimism of the others.

the thing i got wrong at the start: i thought the comparison would end with a winner. it did not. the winner is the task. the model is the tool, and the right tool depends on the job.

for the broader question of how i think about ai tools: what is an llm.