Methodology

Best AI Ever is a database of AI models, providers and list prices, with comparisons, a cost calculator and use-case rankings built on top of it. This page explains where the data comes from, how we check it, what our labels mean and how rankings are calculated. It describes what the software behind the site actually does, so you can judge every number yourself.

The short version:

• Facts and prices come only from the providers' own published pages. Every price links to its source, model pages list the sources behind their facts, and every model and every price carries a verification date.
• If no official source documents something, we show it as unknown. We never estimate.
• Prices are list prices in the currency the provider publishes. We never convert currencies.
• Rankings follow a published formula with visible weights. Missing data never counts as zero, and there is no manual override of positions.
• No one can pay to be listed or to rank higher. The site may show ads, and ads have no influence on data or rankings.

Where the data comes from

Official provider sources only

Model facts and prices come from the provider's own published pages: model documentation, pricing pages, API references, model cards and official announcements. This covers context windows, maximum output, release dates, knowledge cutoffs, supported inputs and outputs, features such as tool calling, API identifiers, licenses and prices. When a provider publishes model weights on a hub such as Hugging Face under its own account, that model card counts as an official source. We do not take facts from resellers, aggregator sites, social media posts or news coverage.

Facts are stored with a link to the page that states them. Our data import refuses a context window, maximum output, release date, knowledge cutoff, license or API identifier that does not cite its source, and no price row can be saved without both a source and a verification date. Model pages list these sources, and each price row links its own.

List prices, as published

Prices are the standard list prices for pay-as-you-go API use, exactly as the provider publishes them. They do not include taxes, negotiated or volume discounts, free tiers, promotional credits or reseller pricing. Where a provider shows a temporary sale price, we record the regular list price and mention the sale price in the row's note. Where a price depends on the region, the note on that row names the region, for example US East for Amazon Bedrock list prices. Subscription plans for consumer chat apps are not part of the price data.

Each price is shown in the currency the provider uses, and we never convert between currencies. As of October 2026, every price in our catalog is published in US dollars.

Every price row has a start date ("Since") and a verification date ("Verified"). When a provider changes a price, we add a new row and close the old one, which stays in the model's price history. When we re-check a price and it has not changed, only its verification date moves.

Unknown means unknown

If no official source documents a value, we record it as unknown. We do not estimate it, infer it from similar models or fill the gap from unofficial sources. On the site, a missing value appears as:

• "Unknown" in capability grids, meaning we could not verify the capability, not that the model lacks it;
• "No data" in comparison and ranking tables;
• a missing line in a model's key specs, or "Unknown" for API and consumer-app availability.

Knowledge cutoffs are shown as month and year (for example "June 2026"). Many providers state only the month, and we do not invent a day.

How facts are verified

Research and an independent second check

Data is collected provider by provider. Before it enters the database, each provider's data goes through a separate, independent second check against the same official sources: the cited pages are re-read and every price, limit, capability and license is compared again. Values that do not hold up are corrected or removed before import.

AI tools do much of this research and checking, and they help draft our texts, including this page. They do not replace sources: a statement without an official source is not accepted as a fact. Imported data always arrives as a draft and nothing is published automatically; an editor decides what goes live. Before an AI-assisted article can be published, an editor has to confirm a review checklist that covers facts, sources, originality and copyright.

Conflicts and open questions

Official pages sometimes disagree, for example a launch announcement and the pricing page. In that case we re-check, generally follow the provider's current pricing page or technical documentation, and record the discrepancy, either as a note on the affected row or as an open question (see below).

Questions that official sources leave open are logged as open questions during research. Typical cases are a model page that does not state a license, or a feature whose support for a specific model no official page confirms. Such values stay unknown until an official source settles them. Limits that a provider publishes only in rounded form are recorded as published, so "1M" becomes 1,000,000 tokens.

Verification dates

Each model shows a "Last verified" date: the day its data was last checked against its sources. Each price row shows its own verification date. These dates drive our freshness rules, described under Updates and review.

License definitions

Each model gets one of three license labels, or shows its license as unknown when no official source states it. The label describes how the model weights are released. Model pages show it together with the name of the license or terms, for example "Open weights (Llama 4 Community License)".

Two examples show where the line runs. The Llama 4 Community License allows commercial use but requires a separate license from Meta above 700 million monthly active users ^[1]. Command A's weights are released under CC-BY-NC, a non-commercial license ^[2]. Both models therefore count as open weights. Apache 2.0 and MIT are on the OSI's list of approved licenses ^[3], and OpenAI publishes gpt-oss-120b under Apache 2.0 ^[4].

A few points to keep in mind:

• The label refers to the weights only. "Open source" on this site does not mean that training data or training code are published. The OSI's Open Source AI Definition asks for those as well ^[5].
• Each license label cites the page that states it: a model card, a download page or the provider's page for an API-only model. If we find no official statement, the license stays unknown. As of October 2026 this applies to two models in our catalog.
• Labels summarize; they are not legal advice. Read the full license before you deploy a model.

Price dimensions

Providers bill for different things, so we store prices per dimension, each with a fixed unit:

Prompt-length tiers

Some providers charge more once a prompt exceeds a certain length. Gemini 3.1 Pro, for example, is listed at $2.00 per 1M input tokens for prompts up to 200K tokens and $4.00 for longer prompts (as of October 2026) ^[6]. We store each tier with the prompt length at which it starts. Tables such as AI pricing show the base tier with a "tiered" badge, and the model page shows every tier.

Listed, never ranked

"Tool and search requests" and "Other" hold charges that cannot be compared fairly across models: per-search fees, priority, flex or off-peak rates, fast modes, explicit cache reads and similar. These rows are listed together with what they cover, but we never sort or rank models by them.

The text-only ranking rule

Wherever text-token prices are compared across models, only models that can produce text take part. This applies to the cost calculator, the text-token tables on the pricing page and ranking criteria based on token prices, including the blended price. Image-only and speech-only models appear in their own dimensions, such as image generation, instead of being ranked by token prices that do not describe their main output. A model whose text output is unknown stays in: for this rule, missing data is no reason to leave a model out.

What the cost calculator includes

The AI cost calculator multiplies your monthly token volumes by each model's list prices for fresh input, cached input (if you enter a cached share) and output, or the batch input and output prices if you choose batch processing. Models documented as not available via API are left out. If a model lists no cached-input price, cached tokens are charged at the input rate and the result says so. Context window and maximum output are only checked when you enter the number of requests, because only then is the average request size known; the same number decides which prompt-length tier applies, otherwise the base tier is assumed and flagged. Models that lack the prices needed for an estimate are listed with the reason. The estimate does not include cache writes, tool or search fees, taxes or free allowances.

Benchmarks

Each benchmark result is stored with the benchmark, the setting it was measured in (for example "pass@1, no tools"), the score, who reported it, when it was measured and the source that published it.

• Self-reported or third party. Results published by the model's own provider are labeled "Self-reported". Results from a benchmark maintainer or another independent party are labeled "Third party". So far we have run no benchmark tests of our own; every score comes from a published source.
• Settings matter. The same benchmark can be run with different prompts, tool access, reasoning effort, numbers of attempts or harness versions, and scores move with them. We compare scores only within the same setting. A comparison table shows a benchmark only when at least two of the compared models have results in the same setting.
• In rankings, a benchmark criterion names one benchmark and either one setting or a setting group. A setting group lists the stored settings we treat as comparable for that ranking, typically the same benchmark version and task set with the same tool access at different reasoning efforts. We never group runs with and without tools, and we leave out subsets and runs whose version is not stated. The methodology box of each ranking lists every setting in its groups, and the ranking table shows the setting behind each score.
• Which result counts. Third-party results are preferred over provider results. If only a provider result exists, it is used and marked "self-reported". Within a setting group, a model's best qualifying result counts; with a single setting, the most recent one counts. Grouping lets us include more and newer models, at the cost of comparing runs that differ in reasoning effort or harness, so small differences between models should not be over-read.
• Caveats. Benchmarks measure narrow tasks. Test questions can leak into training data, providers choose which results they publish, and small differences are often within run-to-run variation. Where we have caveats for a specific benchmark, model pages show them next to the results.

How use-case rankings are calculated

Pages such as Best AI for coding rank models with a formula whose criteria and weights are shown on the page itself. Every ranking follows the same steps.

1. Candidates. Editors choose a list of candidate models and a filter, for example confirmed API access, a minimum context window or required capabilities. A model takes part only if it is published and passes the filter. A capability that is unknown does not pass a filter. In every ranking, whatever its filter, a model must also be available through an API or be labeled open weights or open source (license definitions), so models that can only be used inside an app or a subscription are not ranked. A model whose API access is unknown counts as having none unless it carries one of those labels.
2. Criteria and weights. Each ranking uses criteria of five kinds: benchmark score, price, context window, capability and editorial score. Their weights must add up to 100%.
3. Raw values. Price is a blended price per 1M tokens, calculated from the standard base-tier input and output prices and weighted by the ranking's input-to-output ratio (3:1 unless the page states otherwise). A model listed at $2 input and $8 output per 1M tokens blends to (3 × 2 + 1 × 8) ÷ 4 = $3.50. A price criterion can instead use a single dimension, such as the price per generated image. Context window is the documented window in tokens. Capabilities count yes as 1, partial as 0.5 and no as 0; graded capabilities are spaced evenly, so reasoning counts none as 0, optional as 0.5 and native as 1. Benchmark values are chosen as described above. An editorial score runs from 0 to 10 and must come with a written rationale.
4. Scale. Each criterion uses a linear or a logarithmic scale, as shown in the page's weights table. The logarithmic scale, log10(1 + value), is meant for criteria such as price and context window, so that a few very large values do not flatten the differences among the rest.
5. Normalization. Each criterion is rescaled to 0–100 across all candidates that take part and have data for it. The best value gets 100 and the worst 0, where "best" means lowest for criteria such as price. If all candidates share the same value, each gets 100. Scores are therefore relative: 100 means best among these candidates, not perfect, and adding or removing a candidate can change the scores of others.
6. Score and coverage. A model's score is the weighted average of its normalized values over the criteria for which it has data. Coverage is the share of the total weight that those criteria represent. A model below the ranking's coverage threshold (60% unless the page states otherwise) is listed as "Not ranked: insufficient data" instead of being scored with zeros.
7. Order. Models are ordered by score, then by coverage, then alphabetically. There is no manual position override. Labels such as "Best value" are editorial badges shown next to a model; they do not change its position, and editorial scores count only through their visible weight.

Currency

Price criteria compare prices in one currency: US dollars, unless every candidate is priced in the same other currency. A model priced only in a different currency has no data for that criterion. We do not convert.

Snapshots and "Ranking as of"

A ranking is published as a snapshot. The snapshot stores the criteria, weights, filter and methodology version together with each model's raw and normalized values, the dates of the data and its sources. The page always shows the stored snapshot with its "Ranking as of" date and methodology version, so the text and the table cannot drift apart. A new snapshot replaces the previous one, and any change to the criteria, weights, filter, price ratio or coverage threshold raises the methodology version.

When the data changes

Every day, each published ranking is recalculated in the background from current data. If any input differs from the published snapshot, such as a new price, a new benchmark result or a changed capability, the page is marked "Update in progress". Re-verifying data that has not changed does not count as a change. An editor reviews the differences and publishes a new snapshot; until then, the page keeps showing the last published one.

Comparisons

Model comparisons on the compare page are built from the same structured data: prices per dimension, context window, maximum output, knowledge cutoff, license, release date, availability, capabilities, benchmark results that the compared models share in the same setting, and each model's listed strengths and weaknesses. The coding, writing, research and reasoning rows show each model's rank in the use-case ranking linked to that row, once that ranking is published. Until then, the row shows related data instead, such as the reasoning capability or benchmark results in that area. Wherever a value is missing, the table shows "No data".

Updates and review

• Data freshness. Model data counts as stale 60 days after its last verification, and a price 30 days after its verification. Stale dates are highlighted on model pages. A daily check marks published models with stale model or price data "Update in progress" until an editor has re-verified them against their sources and released them again.
• Price changes are added as new rows, and model pages keep the full price history.
• Rankings are checked every day for changed data, as described above.
• Articles have review intervals, by default 60 days for rankings and comparisons, 180 days for guides and the guides overview, and 365 days for pages such as this one. A published article past its review date is flagged for review; rankings, comparisons, guides and the guides overview then show "Update in progress" until the review is done.

Articles show when they were published or last updated, and model pages show when their data was last verified.

Editorial independence and advertising

• No paid placements. Providers cannot pay to be included, to rank higher or to change how their models are described. Rankings follow the formula above, without manual overrides.
• Advertising may appear. The site may show ads delivered by an ad network, and those ads may promote AI products, including products of providers we cover. Advertisers have no influence on which models we list, what data we show or how models rank.
• Ads stay apart from the data. Ads never appear inside ranking tables or tool controls, and pages such as this one are set up to show no ads. Ad scripts do not load before you have made a choice in our consent dialog; the privacy policy explains the details.

Corrections

If you find an error, please contact us using the details on our imprint page. It helps if you include:

• the address of the page,
• the value you believe is wrong,
• a link to the official provider page that shows the correct value.

We check reports against the official source. If our data was wrong, we correct the record and update its verification date. If the provider has changed a price, we record the new price and keep the earlier one in the price history. We change data only on the basis of official sources.