Independent AI buying guide

No paid rankings

Last checked 26 Oct 2026

Which AIshould I buy?

Methodology

How we rank

Short version: independent tests decide the rankings, a public formula turns them into scores, official pricing pages tell us what each subscription includes, and nobody can pay to change any of it.

The idea in one sentence

Benchmarks test AI models, but you buy a subscription. So we check which models each subscription includes (on the official pricing page), and give every subscription the score of the best model it gives you.

The formula

  1. Each test becomes a score from 0 to 100. 100 is the best model from the companies we cover. For tests with a percentage (like "fixed 80% of the bugs"), a model's score is its share of the best result. For head-to-head votes (Elo ratings), it's how often it would win against the best model, scaled so the best = 100.
  2. Tests are combined per topic with the weights in the table below, averaged over the tests a model has results for. A model is ranked once it has results on at least 2 of a topic's tests, so brand-new models aren't left out while slower tests catch up. Every score shows how many tests it's based on ("based on 2 of 6 tests"): the more tests, the more solid the score.
  3. A subscription gets the score of its best model. If a plan only gives limited use of a model (as most free plans do), that model counts for 90% of its score.
  4. Best value (paid plans only): we subtract 4 points every time the monthly price doubles, starting from €1/$1.
  5. Is the free plan enough? "Yes" when the best free plan scores at least 92%of the best paid one, "mostly" from 80%, otherwise "no".
  6. The 30-second test averages the topics you pick, then adds 4 points if the tool works inside the apps you use, 4 points for a sourced privacy promise (such as "no logs") if privacy matters to you, and up to 3 points for ease of use if you're new to AI. Tools with a privacy warning are left out if privacy matters to you. If you say it must be a European company, the test only picks European options and shows the best non-European one next to it. When no European AI has test scores for what you picked, you get the best European all-rounder, and the test says so.
  7. Stay or switch? If you tell the test what you pay for now, it compares that plan with its pick in the same tests. Within 92% of the pick: stay, unless it costs more than your budget or has a privacy warning when privacy matters to you. Otherwise: switch.

Which tests count for which topic

Only the kinds of tools that make sense compete in a topic: a coding tool isn't "the best AI for writing".

Overall

Coding

Writing

Research & search

Maths & hard problems

Studying

Business & office

Images

Our own opinion: only with a visible reason

Benchmarks miss things like ease of use or customer service. So we can overrule a ranking, but only with a written reason that appears right next to it, marked "Editor's choice". "Our take" texts on tool cards are labeled as opinion and dated.

Not released yet? Never a recommendation

Some models appear in tests before you can buy them. They never become a recommendation. When one beats every subscription in a topic, we list it under "Coming soon", without a price or a buy link.

What counts as a source

Every benchmark source has to meet four requirements:

  1. Independent of the AI companies it measures.
  2. Open about its method, so anyone can check how scores are made.
  3. Actively maintained. Our update program warns us when a test hasn't added new models for six months, and we drop it.
  4. A licence that allows reuse, with credit where the licence asks for it.

For prices and plan contents we only use the company's own pricing page. Pages that block automated visits are checked by hand and marked until then.

How we keep it fresh

Every source we use

Independent tests

Official pricing pages

Leaderboard data: Arena (arena.ai), CC BY 4.0 · Benchmark data: Artificial Analysis (artificialanalysis.ai) · Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks, CC BY 4.0 · Benchmark data: LiveBench (livebench.ai)