How AI Compare works

What we compare

Individual plans of AI assistants, research products and coding tools, as one person can buy them in United States: the free tier and the individual paid tiers. Team, business and API prices are kept apart and never fill an individual plan's field. A product that gives access to another company's model does not inherit that company's own app features. Choosing a model for your own software (APIs and token prices) needs workload-based costs and its own tests, and is not part of AI Compare yet.

What a fact is

Every fact is one plan and one capability, with a state (available, not available, unknown, not applicable), the conditions that qualify it (limits, surfaces, region, prerequisites, gradual rollouts), the kind of evidence, its source, and its dates.

• Vendor documentation states: an official pricing page, help article or announcement says so.
• WhatTheTech observed: we ran a test kit ourselves; the run, the plan, the date and the result are recorded.
• A third-party evaluation measured: someone else's measurement, within its own documented scope.
• A personal observation: what you record in a scorecard. It stays in your browser and is never published.

Documented access is not performance. "Supports file uploads" says nothing about how well a spreadsheet is analysed, and a training opt-out says nothing about how long data is kept. Quality claims need test evidence, and we show none we did not collect.

How facts are checked

An editor verifies each fact against its source and records the date. A fact that is edited becomes unverified again and leaves the public data until it is checked. Facts are published in numbered data releases; every page, the tools and the downloads read the same release.

The initial release (October 12, 2026) was compiled from official vendor pages with AI-assisted research and is marked that way on every fact. Editors re-verify it on the schedule below.

Every fact passes three steps before it is published: draft, source checked (someone confirmed the source says exactly this) and reviewed. A release is a fixed snapshot: a later edit or a retraction changes the next release, never one already published, and a retraction is listed in the change log with its reason. Models are recorded the way each plan's own page names them: a version when the page gives one, the family when it does not, and "automatic choice" when the product picks the model without saying which.

Freshness

• Prices and access that decide a choice: reviewed every 14 days.
• Other decision facts: every 30 days.
• Background descriptions: every 90 days.

A fact past its review date is shown as due. One full interval later it is stale: still visible with its date, but treated as not verified. A stale price is never used to name the cheapest option.

How a recommendation is made (rules 2.1.0)

"Find my AI" asks what you do, how much each task matters, what you cannot do without, your budget, what you already pay for and whether you want one app, one subscription or a few. One calculation in your browser does the rest; the same code produces the cards, the tables, the charts and your saved record.

Tasks and weights

• Each task you choose gets a weight from its importance: Essential counts 3, Very important 2 and Occasionally useful 1. These are our published defaults, not measured values. You can set exact percentages instead; we scale any numbers to add up to 100%.
• Some tasks test much the same skill (editing an email and editing an article, for example). When you choose both, they share one weight, so the same evidence never counts twice.
• Importance changes the order of options. It never removes one, unless you explicitly ask for your Essential tasks to be treated as requirements.

Requirements: pass, fail or unknown

• Every plan, and every pair of plans when you accept two tools, is checked against your must-haves, your budget on the billing basis you chose (monthly payments unless you allow annual plans) and any plan it depends on.
• A requirement the plan documents as missing is a fail, and nothing else the plan offers makes up for it. A requirement that is not documented, out of date or contested is unknown: the option is shown separately, to check first, and never described as meeting your requirements.
• For "every tool in a combination" requirements (no ads, a training opt-out) each plan in a pair must meet it.
• A pair that includes a plan documenting nothing you asked for is not built: it costs more than the other plan alone and covers the same, so it could never rank above it. Plans you already pay for and keep are always included.

Documented coverage

A task is covered when one plan documents every feature the task needs (a research task needs web search with sources, for example). Coverage is the sum of the weights of the tasks covered. It says what the plans offer, according to their own pages. It is not a quality score, and tasks whose features are not verified count as not covered while being shown as "not verified", never as "not offered".

The result card says how many of your tasks have every feature they need documented ("3 of 4 tasks"), and that answer quality is not tested. The weighted coverage percentage is in the details and the record, with this definition.

Options are ordered by coverage, then fewer subscriptions, then fewer apps, then lower monthly cost. Plans with the same coverage, setup and cost are shown as a tie: the facts cannot separate them, and a test can. Beside the leading option we show:

• the cheaper option with the most coverage, labelled "Same coverage, lower cost" when it documents as much as the leader (it can be a pair of plans, or a cheaper single plan);
• a pair that documents more than the best single plan, when you accept two subscriptions;
• the best single plan, when the leader is a pair.

Fewer subscriptions come first only among options with the same coverage; a cheaper option with the same coverage is never hidden. A pair is scored by giving each task to the plan that documents it, which shows coverage, not that two tools work well together.

Apps and subscriptions

A subscription can include more than one app: The X Subscription includes Cursor and Grok, ChatGPT plans include Codex, Claude's paid plans include Claude Code, and Google AI plans include Jules, Google Antigravity and Gemini Notebook. Each fact records the app that delivers it, so a result counts subscriptions (bills) and apps (what you open) separately. An app counts on its own when its maker sells or names it as its own product. Add-ins, extensions and features inside another product, such as ChatGPT for Word, Claude in Chrome or Gemini in Gmail, count under the assistant they belong to. Choose "One app" and an option whose tasks run in two apps is excluded; choose "One subscription" and it is allowed, with each app named.

Your goal in your own words, and projects

You can describe what you want to do instead of picking cards. Your words are matched in your browser against a fixed list of phrases for each task; no AI reads them and nothing is sent or saved. The tasks we understood are shown as ticks for you to correct before anything is calculated. A job no product in our catalog covers yet (presentation makers, translation or scheduling, for example) is named as not covered, with the closest task we do compare offered, never quietly put in its place.

A project (a music video, an article, an app, product photos, a presentation, a spreadsheet review, meetings) is a series of stages. Stages that must happen count as requirements and optional ones weigh little; a stage you already have, such as an existing song, leaves the plan, so you are never pointed to a subscription for it. Each stage shows the app and plan that documents it. Between stages we say only what our data supports: the same app, where the work stays put; a manual transfer (you move the file yourself; we have no tested or documented hand-off between those apps) or unresolved (no product in the catalog does the next stage). Two apps that open the same file type are never presented as working well together.

Measured performance

Our test kits are the evaluation protocols. A plan's result on a kit is the share of the kit's checks passed, averaged over our approved runs of the current kit version. A result counts only with at least 3 runs, made after the plan's last material change (a new model, new limits or a renamed tier ends earlier runs), and no older than a window that depends on how fast that kind of product changes: research 90 days , writing 180 days , coding 120 days , images 120 days , video 90 days and data 180 days . These windows are our editorial judgement, not a statistical finding: search, research and video products change fastest, writing and spreadsheet work slowest. Options are ranked on performance only when every option compared has a qualifying result for every task you chose; the ranking then covers only tested plans, and the others stay listed as "not enough comparable evidence". Two results closer than one check of the kit (10 points on a ten-check kit, 20 on a five-check kit; over several tasks, the same band weighted like the score) are "too close to distinguish": the kit cannot tell them apart, so neither can we. This is a practical rule tied to the kit's size, not a confidence interval. Missing results are never scored as zero, and weights are never re-spread to hide a missing task. Until results are published, every order is by documented coverage, and the page says so.

There are two kinds of kit. A fixed-file kit gives every AI the same frozen files, so runs on different days can be compared; only these count towards results. An open-web kit asks a current question whose right answer changes over time; it records the date and the search settings, its runs are compared only within the same day, and it never counts towards a ranking. Coding kits come with held-out tests that the reader keeps back from the AI: they show whether code follows the task or only the tests it was shown. Code is run on the reader's own computer, never on this site.

Readers can score answers blind on a kit page: they paste each AI's answer, score them under neutral labels and reveal the names at the end. The answers are not saved; the scores stay in the reader's browser. A reader who ran the tests may still recognise an answer, so blind labels reduce that bias without removing it, and a personal scorecard is an observation, not a public benchmark.

What would change the result

• For each task we try one step more and one step less importance, and say which change would move the leading option.
• We compute the weight a task would need for another option to draw level, with the others scaled down in proportion.
• We relax one condition at a time (annual billing, a second tool, the budget, one must-have) and say what that would allow. Nothing is ever relaxed without you choosing it.
• When you change an importance on the results, the page shows each task's contribution before and after; any other change you made is listed separately.

Each result names the data release, the rule version and a calculation number, and lists the facts it rests on. A saved decision keeps its original result; reopening it recalculates on today's data and shows what changed.

Money

All amounts are whole cents. The monthly equivalent of a yearly plan is the yearly price divided by twelve, but a yearly plan is paid up front and is not cancel-any-time: dropping it saves money from the next renewal, not today. Amounts in different currencies are never added together. A missing or unverified price leaves a total incomplete, and an incomplete total cannot be the cheapest. Sales tax is not included unless a plan says so.

API costs

API prices are kept apart from subscriptions: they are charged per token, so they cannot be set side by side with a monthly plan. We record each company's standard pay-as-you-go list price per million tokens, exactly as its pricing page states it, with long-context prices and announced price changes where the page lists them. A month's cost is calculated for a workload you enter: every input token is billed once, as plain input, cached input or cache write; a request over a model's long-context limit pays the higher prices; and money is counted in exact units and rounded to cents once, at the end. A price the company does not list is never guessed: the estimate says it is incomplete. Batch, flex and priority tiers, regional surcharges, tool fees such as web search, cache storage and volume discounts are not included, and the same text counts as a different number of tokens with each company. No quality, speed or reliability is measured on this page yet.

Your data stays with you

The tools run in your browser. Your answers, subscriptions, scorecards and saved decisions are kept in this browser's storage only when you save them, and can be exported, imported and deleted. Nothing you enter is sent to our server, placed in a link or used for anything else. If your browser blocks storage, the tools still work; they just cannot remember.

To know which parts are used, we count a few actions as anonymous daily totals: a page opened, the finder started and a result shown (by kind of result, not its content), a kit downloaded, a comparison made. No cookie, no IP address and no identifier is kept, and nothing you type or choose for privacy is counted. If your browser sends Do Not Track or Global Privacy Control, nothing is counted at all.

Independence

No result here is paid for. Products cannot buy a better position, and no affiliate link is used in any comparison or recommendation.

Limits

• Vendors change plans often, and some limits are not published.
• Availability can differ by country, app store, account type and gradual rollout.
• We do not run tests automatically. A test result appears only when an editor ran it.
• Free plans and plans described as "limited" or "expanded" do not confirm enough capacity for heavy daily use.
• Feature coverage is about what is offered. Answer quality, speed and reliability need tests, which are listed with every result.