Choose the right AI for the task

Start with the task, not the leaderboard

Every few weeks a new model claims the top of a benchmark. For most people that changes very little. The major providers all offer capable models now, and the differences that matter day to day are price, limits, privacy, what the tool connects to, and how it handles the specific job you have.

So the useful question isn't "which AI is best?" It's "which AI is good enough for this task, at a cost and risk I'm comfortable with?" Five questions get you there.

The five questions

1. 1What exactly is the task?Write it as one sentence: "Summarise supplier contracts", "Draft social posts from my videos", "Fix bugs in a Python project". If you can't, the tool choice isn't your problem yet.
2. 2What does a mistake cost?A weak brainstorm costs nothing. A wrong figure in a client report costs a lot. Higher stakes justify a more capable tier and always justify checking the output yourself.
3. 3How sensitive is the data?Read the provider's data and training settings before pasting anything confidential. Business plans often come with different data terms from personal ones. If data must never leave your own machine, look at open-weight models you run yourself.
4. 4How often, and how much?A few chats a day suits a subscription. Thousands of documents, or building AI into your own product, suits pay-per-use access through an API, where cheaper tiers make a big difference.
5. 5What's the budget?Free plans are fine for learning. A personal subscription costs roughly the same as a streaming service. API costs scale with use, so estimate before you build.

Apps or API: two ways to pay

Pick a tier, then move up or down

Each provider sells a range: a fast, cheap tier for high volume, a balanced middle, and a top tier for long, demanding work. Prices across one lineup can differ by a factor of fifty or more, so the tier matters more than the brand.

• Start with the provider's balanced or recommended model. Anthropic, for example, tells developers who aren't sure to start with Claude Opus 5.5.
• Move down a tier for repetitive, well-defined jobs such as sorting, extracting or tagging, once you've checked the cheaper model handles your examples.
• Move up only when the output from the middle tier isn't good enough on your own work, not because of a headline.

The current options

This list is built from our model profiles, so it shows the same checked facts and dates. Follow a family's name for the full profile, including who each tier is for.

When a model on your own computer makes sense

Open-weight models such as Meta's Llama can be downloaded and run on your own hardware, with tools like Ollama. Nothing you type leaves the machine, it works offline, and there's no per-token bill.

The trade-off is that you supply the hardware, the setup and the updates, and the models that run comfortably on a laptop are usually smaller than the hosted flagships. It suits privacy-critical work, experiments and learning how AI works. Most people should start with a hosted app and add a local model later if they need one. Each Llama version has its own licence, so read it before using one for business.

Settle it with a twenty-minute test

If you're torn between two options, don't read more reviews. Run both on a real task of yours and score the results. Free plans are usually enough for this.

Limits of this guide

• We haven't yet tested these models against each other, so we don't say which writes, codes or reasons best. When we do, results will appear on the model profiles, labelled as our tests.
• Prices and lineups change often. Every figure here has a checked date; look at the provider's page before you commit money.
• App plan prices shown are the providers' US prices. Prices in other countries and currencies can differ.