Claude wins the vibes test. Five anecdotes still aren't a benchmark.

The strongest fair version of the claim: for five common everyday tasks — brainstorming, explaining complex topics, workflow and productivity advice, writing and editing support, and personalized recommendations — Claude Opus 5 produces answers that feel more thoughtful, better structured and more human, and it needs fewer follow-up prompts to get there.

The author is careful in two ways worth crediting. She says she still uses ChatGPT, and she acknowledges that ChatGPT knows her better through memory — which means she is aware that part of the comparison is not like-for-like.

It demonstrated a consistent personal preference across five task categories, described in enough detail that a reader can try the same thing. That is legitimate qualitative evidence and it is more useful than most benchmark commentary, because it reports the thing benchmarks systematically miss: friction.

What it does not contain is any of the machinery that would make it a comparison:

• no disclosed same-prompt, blind head-to-head
• no repeated trials, so run-to-run variance is invisible
• no scoring rubric — "more thoughtful" is never operationalised
• no second evaluator, so preference and expectation can't be separated
• an explicit confound the author names herself: one assistant has memory of her, the other does not

The counter-case is not "actually ChatGPT wins." It is that this genre of article repeatedly produces opposite conclusions from equally sincere writers, which is what you expect when the measurement instrument is one person's taste over a handful of sessions.

Nor can OpenAI settle it. A vendor's own evaluation is testimony, not a ruling — that is rule two of our method. On a subjective preference claim, the only party who can adjudicate is you, running the same prompt yourself.

The most interesting evidence here undercuts the premise of the whole genre. Research on how users evaluate AI chat assistants (Beyond Benchmarks) finds that active users frequently use several platforms rather than settling on one, and that what drives adoption differs from product to product rather than reducing to a single quality ranking.

We should not overstate that paper: it describes user behaviour and stated reasons, not model capability. But it supports a modest conclusion — the "which assistant wins" frame is a poor description of how people actually use these tools.

Plausible personal preference; weak universal evidence. Nothing in the article is dishonest and its observations may well replicate. But "five prompts that replace ChatGPT" is a capability claim resting on experience evidence, and the article does not close that gap.

The genuinely valuable insight is buried under the versus framing: users increasingly route different tasks to different models, and subjective qualities — tone, structure, memory, and how many follow-ups it takes to get something usable — matter alongside benchmark scores.

Don't take our ruling either. This case converts cleanly into a transparent reader challenge: paste the identical prompt into both assistants and judge on the criteria you personally care about.

1. Brainstorming. Ask both assistants for ten angles on the same idea. Judge on range, not polish: how many options are genuinely distinct?
2. Explaining a complex topic. Pick something you already understand well. The right test is whether the explanation is correct, not whether it sounds confident.
3. Workflow and productivity advice. Give both the same messy week. Count the follow-up messages you needed before the answer was usable.
4. Writing and editing support. Paste the same draft. Judge whether the edit preserved your voice or replaced it with the model's.
5. Personalized recommendations. Note which assistant already knows your context. Memory is a real advantage, but it's an account effect, not a model capability.

One methodological note that costs nothing: run each prompt twice per assistant. If the two runs from the same model disagree more than the two models disagree with each other, you have measured variance, not quality.

If accuracy rather than feel is what you're weighing, read Case 003 on AI error rates, which takes apart the numbers people cite in these arguments.