AI Chat & Assistants

What independent reporting tells us

The central weakness is no longer simply whether an assistant can produce fluent text. It is whether users can trust it with a real task. In WIRED's hands-on account of OpenAI's Dots agent, the reporter described useful shopping assistance alongside identity confusion and failures on web obstacles. That experience concerns an agentic feature, not every ChatGPT conversation. It nevertheless illustrates why an assistant's ability to take action needs closer scrutiny than its ability to write an answer.

The Verge's reporting on Google's enterprise Gemini agent likewise highlights the growing importance of connections to workplace apps. Access is not the same as reliability: a useful evaluation should check what the assistant can read, what it can change, and what requires approval.

Four tests before you subscribe

First, give two candidates the same real assignment. For research, ask for three primary sources and check every link. For writing, provide a rough draft and compare factual preservation, not just polish. For coding, run the resulting tests. For analysis, verify one numerical result independently.

Second, measure correction effort. A fast answer that takes twenty minutes to repair is worse than a slower one you can trust. Record mistakes, omitted instructions and how well each tool handles feedback.

Third, inspect data boundaries. Consumer chats, team workspaces and API products can have different retention, training and administrative controls. Read the policy for the exact plan you intend to use.

Fourth, compare actual limits: messages, file sizes, advanced model access, tool usage and seats. Headline monthly prices conceal much of the difference. Check each provider's current pricing page before buying.

Belliste editorial verdict

Start with ChatGPT if you want one flexible assistant; compare Claude for document-heavy writing or development, Gemini for Google-centric work, and Perplexity for source-led discovery. Those are workflow recommendations, not claims that one model wins every benchmark. The sensible choice is the tool whose errors you can detect and whose limits fit your working week.

For detailed specifications, open the individual Belliste model pages and compare two assistants side by side.