Guide

ChatGPT, Claude or Gemini? Test them yourself in an hour

Leaderboards won’t tell you which AI assistant does your job better. A blind test on five real tasks will. Step by step, with a scoring sheet.

Published 3 min read Accurate as of November 21, 2026

Verdict

There is no single winner for everyone. One hour of blind testing on five tasks from your own work tells you more than any leaderboard.

In the autumn of 2026 all three companies replaced their strongest models within a few weeks of each other. Leaderboards reshuffle constantly, and benchmarks measure tasks you will probably never do. The good news: you don’t need benchmarks to pick an assistant for your work. You need an hour and five real tasks.

Who’s in the race (as of 21 November 2026)

Assistant Company Current model family
ChatGPT OpenAI GPT-6 (plus GPT-6.2 Astra and Sol in the API since 17 November)
Claude Anthropic Claude 5.5: Opus, Sonnet and Haiku, plus the top-tier Fable 5.2
Gemini Google Gemini 4 (including Argon and Neon)

Version names change fast, so we update this table with every major release. The test below works whichever version happens to be newest.

Step 1. Pick five tasks from your actual job

Skip riddles and trivia. Use things you really do every week, for example:

  1. a reply to a difficult client email,
  2. a summary of a long document or report,
  3. a piece of marketing copy or a social media post,
  4. an analysis of a small table of numbers,
  5. one task from your field where mistakes are easy to make.

Prepare the material in advance and strip out personal data and anything confidential.

Step 2. Same prompt, clean start

Give every assistant the identical instruction and the same files. Open a new chat and switch off memory or personalisation for the test if you use it. Otherwise you are comparing what each one already knows about you, not the models.

Test the plan you will actually pay for. If you plan to stay on the free tier, test the free tier.

Step 3. Score blind

This is the step that matters most. Paste the answers into one document, label them A, B and C, and keep the key somewhere else. People go easier on the tool they already like; blind scoring removes that.

Step 4. Rate each answer on four points

For each task give 1 to 5 points for:

  • accuracy – no factual mistakes,
  • ready to use – could you send or paste it as is,
  • edits needed – how much you had to add or change,
  • tone and style – does it sound like you and your company.

Five tasks, four criteria, three assistants: 60 scores. It sounds like a lot but takes about 20 minutes.

Step 5. Check facts and figures

For number tasks, work the result out yourself or in a spreadsheet. For summaries, check that nothing appears that wasn’t in the document. One invented fact in a client email costs more than the price difference between subscriptions.

Step 6. Add price and convenience

Only at the end look at cost and where you work. If you spend your day in Google Workspace or Microsoft 365, the integration may be worth more than a few points in the test. Check current prices on the vendors’ own pages – they change often.

Scoring sheet to copy

Task A B C Notes
Client email /20 /20 /20
Summary /20 /20 /20
Marketing copy /20 /20 /20
Number analysis /20 /20 /20
Field-specific task /20 /20 /20
Total /100 /100 /100

Common mistakes in tests like this

  • One task instead of five. Assistants differ by type of work. One task is chance; five give you a picture.
  • Different prompts. Even a small wording change can move the result more than switching models.
  • Scoring by gut feeling. Without points, the one that writes most nicely usually wins, not the one that is most correct.

Most often one assistant wins at writing and another at analysis. Then the sensible move is to use two – and know which one to reach for when.

Sources

  1. OpenAI – Model Release Notes
  2. Anthropic – Models overview
  3. Google – Gemini 4 Argon: our next era of frontier intelligence
  4. Google – Introducing Gemini 4 Neon