Home · Insights · Fundamentals
ChatGPT vs Claude vs Gemini: how to choose for your business
Rankings go stale in months. A method does not. Here is how to choose between the big three assistants for a small business in three steps, and why you should build so the choice can change.
The honest answer is a method, not a winner
Which assistant is best changes every few months, and comparisons written for this year go stale before you finish reading them. What lasts is how to decide for your business. We do three things: check what each provider does with your data, test them on your own work, and build so the choice can change.
1. Read the data terms before the feature list
For a European business the first question is not which model writes best. It is what happens to what you paste in. The answer depends on the plan, and the consumer and business versions of the same product often differ. Read the terms for the exact plan you will pay for, and check three things: whether your content is used to train models, how long it is kept, and where it is processed.
We have read the published terms for Anthropic commercial terms and Google Cloud data processing terms. Anthropic's commercial terms state that it may not train models on customer content from its services. Google's Cloud addendum covers Google Cloud services and sets out how it processes customer data on the customer's instructions. Neither is a statement about the free consumer apps, which have different terms. We have not quoted OpenAI's here, so check theirs directly. None of this is legal advice, and if you handle health, financial or other sensitive data you should have someone qualified read the contract. For the wider picture see data privacy and AI for a European business.
2. Run a bake-off on your own work
Public benchmarks measure what the vendor chose to measure. Your own work is the only test that counts, and it takes an afternoon.
- Pick five real tasks. A customer email, a summary of a long document, a spreadsheet question, a first draft of a proposal, a messy note to be turned into a checklist.
- Write what good looks like first. Three criteria per task, for example accurate, correct tone, no invented facts. Decide before you look at any output.
- Give each tool the same input, with no extra tuning.
- Score blind. Have someone strip the tool names, then score against your criteria.
- Note the failures, not just the best answers. A tool that is brilliant most of the time and confidently wrong the rest is the riskier of the two. We explain why in the jagged frontier.
3. Do not marry the model
| Choose by | Why it matters |
|---|---|
| Data terms for your plan | Decides whether you can use it with customer information at all |
| Results on your five tasks | Beats any public ranking for your purposes |
| How easily you can switch | Models change fast. A system built so the model is one swappable step costs little to change |
| Price at your volume | Light users are fine on any plan. High volume through an automation is where costs differ |
We build the AI step in our automations as a single, replaceable component. When a better or cheaper model arrives, we change one step and test it against the same examples. That is why we are comfortable recommending a model today and telling you it may not be the one in a year.
Where we land
We use Claude for most of our own work and have said so in the AI stack we run Viacala on. That is our result on our tasks, not a ruling for yours. If you want help running the bake-off or choosing the plan, it is part of the AI tools we use on Mallorca client projects.
Sources, checked 9 October 2026
- Anthropic, Commercial Terms of Service, read as published.
- Google Cloud, Data Processing Addendum, read as published.