How we score.

A review tells you whether something new is worth using now, compared with what else exists, for what it costs. Each job it competes in gets a score from 0 to 100, and the page names the public sources it is built from. We never run or test anything ourselves.

A score per job

We score each release only in the jobs it really competes in, usually one to four of these: writing, research, coding, data, images, voice, video, agents. 100 is the best available today in that job, and every score is against what exists now, so scores move as new releases come out.

Where something has scores in several jobs, its overall score is the average of them, always shown with how many jobs it covers, such as "92 across 4 jobs". A model can lead in research and trail in code, so the job scores matter more than the average.

What a model score is built from

Model scores are our own. We build them from public data that other people collect, and combine it by our own method, which we don't publish. The sources today:

Blind votes from Arena's open leaderboard dataset, where people compare two anonymous answers and pick the better one.
Independent tests from Epoch AI, which runs well-known tests on models itself. We only use tests Epoch ran, never a maker's own results.

Each review names the sources behind each of its scores. The likely range next to a score, such as "84, likely 81 to 87", shows how far it could be off given the evidence so far. When two ranges overlap, we call it a tie rather than pretend one is clearly better.

Agents. Arena's blind votes for agents give a rank but no rating, so models there show their rank, such as "#3", and no score yet.

Updated weekly. We take a fresh copy of each source every week; last read 20 October 2026. Arena leaderboard data © Arena, CC BY 4.0. Epoch AI benchmark data, CC BY 4.0.

Scores from our checklist

Apps and tools are scored on our own checklist, which looks at five things:

Does the job
Quality underneath
Value
Privacy and control
Limits and reliability

What gets no score

First looks. On release day, before there is enough evidence, a release gets a First look: one of three calls, Worth trying, Wait or Skip, from the maker's own pages, and no score.
Voice. None of our sources covers voice yet, so voice models and apps show "No score yet", with the evidence we have written out in words.

Makers' own results on benchmarks are shown beside a score as context, labelled with who ran them. They never go into the number.

The call

Every review ends in one of these five calls. A First look uses only the three marked "First look too": on release day there is not enough evidence to say Use it or Depends.

Use it	Among the top scores in its best job, and good value. A pick we would point a newcomer to.
Worth trying	Promising, or a strong pick for a specific need: cheaper or more private than the leaders. First look too.
Depends	Good at one thing and weak at another people will expect. The line says which.
Wait	Announced or early: a waitlist, a preview, missing evidence, or likely to change within weeks. First look too.
Skip	Outclassed or not worth its price. The line names the better alternative. First look too.

What gets a review

We focus on the best things rather than flooding the list with low scores. A release gets a page only if it is one of these:

Big news: a launch many people will hear about this week. These get an honest score even when it is low, so you know it is noise and what to use instead.
Very good value: a strong score in a job and a Best buy or Best free there, or clearly cheaper than the current pick at about the same score.
On a Top 20 list: any model in the top 20 of one of our lists; apps and tools only when they make the top 20 of theirs.

A Best buy is on the frontier for its job: nothing we have scored is both cheaper and higher. Best free is the best score at $0, from a free plan or free local use. Best on your computer is the best score among things you can run on your own computer, where nothing leaves your machine.

Rules we don't break

No sponsors, paid placement, affiliate links, or free access in exchange for coverage.
We never run, test or measure anything ourselves, and never say or imply we did.
Every figure links its source and the day we opened it.
Makers' claims are attributed to them and never decide a score.
Scores are against what exists today, so they go stale. Our sources are read again every week; we re-check every review within 30 days and log every change on the page.
Products from every maker get the same method: each score is built from the same dated data, the same way for everyone.

Spot a mistake?

Every review page has a Report a mistake link that opens an email to us. If you ever see a sentence that sounds like first-hand testing, tell us. Corrections are dated, and each one shows in News marked as a reader report.

Who drafts the reviews

Reviews are drafted with help from AI models, including Anthropic's Claude. Model scores come only from outside leaderboards and independent evaluations, by a written rule. Where a rule doesn't settle a case, we take the lower score and say why. Reviews of products from a maker whose models help draft our reviews (today, Anthropic) say so at the top of the page.

Download the data

Everything on these pages is also a file you can use, rebuilt every time a review changes. Our scores, verdicts and write-ups are free to reuse under CC BY 4.0: credit Hayz Labs (hayzlabs.com/reviews).

data.json: every review: a score from 0 to 100 per job with its likely range and the public sources it is built from, verdict, price and the dated sources behind each.
list.csv: All releases, with dated prices and facts and each one's arena.ai places.

Values from others are not ours to license. Our model scores are built from public data, including leaderboards Arena publishes as an open dataset under CC BY 4.0 (credit "Arena leaderboard data © Arena, CC BY 4.0") and Epoch AI's benchmark results under CC BY 4.0 (credit "Epoch AI"). The 0 to 100 scores are our changes to them. Prices and specs are makers' published facts, linked on each row. We leave Artificial Analysis figures out of these files because its terms do not allow bulk reuse; review pages cite them with a link.