How it works

Four steps, shown for coding and building apps, the task people search for most.

1. 1

   Gather the evidence

   Tests and votes that measure the task, each labelled by how closely it matches.

   3 tests for coding and building apps

2. 2

   Score each model

   Combine its place on every test into one task score.

   Claude Fable 5.1 comes 1st of 10

3. 3

   Make the picks

   Best overall, best for $25 a month or less, and best free.

   3 picks

4. 4

   Match to a plan

   Find the cheapest app and plan that includes each pick.

   Claude Max, $100/month

Our own tests, plus free, public sources whose licences let us show their data. Every chart names its source underneath, and our own results carry the navy Tested by us tag.

Hardly any test measures an everyday task exactly, so every test on a task page carries one of three labels.

A more specific task, like vibe coding, borrows its broader task's tests when it has none of its own, labelled as related at most.

On each test, the lowest of our models is placed at 0 and the highest at 100. A model's task score is the average of its places. Here's how the top of coding and building apps is worked out:

• It's relative. It shows who leads among the models we track, not how good they are in absolute terms. Last place isn't useless.
• A test needs at least 3 of our models before it counts, so two results can't decide a ranking.
• A model needs results on at least half the tests, including at least one test of real work when the task has one. People's votes alone can't put a model on the chart.
• A dash means the model hasn't taken that test; its score averages the tests it has.

4 more models have some results for coding and building apps but not enough to score yet.

Every task page opens with picks made by a fixed rule, so they always agree with the charts:

• Best overall: the highest task score, and the cheapest plan that includes it.
• Best for $25 a month or less: the highest-scoring model on a plan at that price or below.
• Best free: the highest-scoring model on a free plan.
• Within 5 points: it's too close to call, so we name both, higher-ranked first, each with its plan and price.
• Extra cost doesn't count. A model you pay for on top of a plan isn't treated as included.

Here's the result for coding and building apps:

Results come with a likely range, drawn as a thin line on each bar. When the top two lines overlap, we say it's too close to call rather than naming a winner.

Independent testers take one to three weeks to catch up with a launch, so each model shows where its evidence stands. A new version always starts from nothing; we never assume it performs like the one before.

ChatGPT, Claude and Gemini are apps; each runs one or more models, and free plans often run an older one. We compare the models, then map each to the plans that include it.

• We check plans against the makers' own pages at least every 30 days. Where a maker's page can't be read, we use press coverage and say so.
• An older model that a plan still uses stays on our charts, because it's what people on that plan actually get.
• Plans change often, so check the price before you buy.

Cost charts show what developers pay to use a model directly, for 1,000 typical requests of about 1,500 words in, 500 out. If you use an app, the monthly plan is what you pay.

Norm, our guide

Anything Norm says about the data, like who leads or what the top models cost, is filled in from the same numbers as the charts, so it can't disagree with them. His advice is written by a person and only appears once it's been reviewed.

What the makers say

Launch claims appear on model pages, always labelled as the maker's own, and never count towards a score.

• How good AI is at your exact job. Tests are close stand-ins, and the label on each chart tells you how close.
• Whether one answer is right. Every model makes mistakes; Norm's guide on each task says what to check.
• How a model does inside a specific app feature, like a document upload or a coding tool. We compare the models, not the apps.