With the November election approaching, I gave ChatGPT, Claude, and Gemini the same research assignment. I wanted a CSV listing the races and candidates on my ballot, with a short summary of each candidate's platform drawn from the candidate's own website or social media.
I expected the summaries to differ. What surprised me was that only two of the three assistants completed the assignment.
The prompt
I used Pro-level accounts and left each tool on its default selected model. Each received one identical prompt. There were no follow-up prompts. Here is the exact wording, with the residential address removed:
Generate a csv list of every position my friend will be allowed to vote for based on address <REMOVED> in the 2026 November election. For each candidate review the candidate's official website of the candidate or social media presence if there is no website. For each candidate, summarize in one sentence what that person is running on from a platform priority standpoint. List no more than 3 items in the single sentence summary.The address I submitted was my own. I referred to “my friend” because I didn't want the assistant to associate me personally with that address.
I wanted to see what each tool would do with that request on its first attempt. Going directly to the candidates' own material also mattered to me. I wanted their stated priorities rather than a news outlet's characterization of them. Their own descriptions would still reflect their chosen positioning, of course.
Two completed it
ChatGPT, using GPT-6 at the Medium setting, produced the requested result without a series of interim website-access approvals. From my perspective, it performed best in this exercise because it completed the assignment in one shot with the least interruption.
Claude, using Opus 5.5, also produced a useful result. It required many interim authorizations to read websites along the way. I considered the outcome good, but getting there required more of my attention.
The two outputs had different columns. Some candidate summaries were nearly the same, while others differed more noticeably. That was the kind of variation I had expected to explore.
Gemini redirected me
Gemini responded with voter-resource websites and instructions for finding the ballot information myself. Those resources could help someone do the research, but the requested deliverable was a CSV. Gemini did not produce it.
When I returned to the original, undisturbed conversation, its model selector showed Gemini 3.1 with Advanced Reasoning. The experiment had taken place after Gemini 4 Argon's release, but that did not mean I had used Argon in the original comparison.
I then ran the same prompt using Gemini 4 Argon. It also redirected me to voter resources instead of generating the requested CSV. That was a separate check after the original three-tool exercise, and it produced the same practical disappointment.
Completion matters
There is a substantial difference between receiving the requested work product and receiving directions for doing the work yourself. When I ask an assistant to generate a file, that difference affects how useful the interaction is.
The interruptions also mattered. Both ChatGPT and Claude delivered results I could work with, but Claude required more participation during execution. That is part of the experience of using these products, even though the number of visible approval requests tells me little by itself about the quality of their research.
Completion is only the first question. I have not established here that either generated list was exhaustive or that its candidate summaries were accurate. A finished CSV still needs checking against the official ballot and the underlying candidate material before someone relies on it.
This was one practical exercise using paid accounts and their default selections, followed by an additional Argon attempt. It gives me a specific experience to judge, not a general ranking of these models.
I went into it expecting differences in how three AI assistants summarized the same subject. The larger difference appeared earlier: whether they would produce the work I had asked for at all.