Skip to content
✦
English
Menu

AI MODEL LAB

Research methodology

Model profiles

We prioritize official provider documentation and record version, source URL and verification date. Unknown fields remain unverified; we do not infer technical specifications from names or marketing language.

API pricing

Each rate records platform, mode, resolution, currency, billing unit, unit price, source and verification date. Numeric sorting stays within one model category, billing unit and currency. Modes and specifications may still differ, so the sorted rates are not full task costs. Expired or unverified records are not presented as current prices.

Strength scores: objective and subjective evidence

Objective measures can include task completion, error rate, end-to-end time and generation failure rate under shared conditions. Full cost is also objective data, but remains in the separate pricing comparison and does not affect strength. Text tasks may have verifiable answers; speech can have pronunciation checks. All figures need task, language, model version, settings and provider recorded.

Automated video consistency and speech transcription error are repeatable proxy measures, not substitutes for perceived quality or real-world intelligibility. Subjective evidence includes blinded human preference and judgments of visual or audio quality. User feedback and test videos may contain both preferences and verifiable facts; we label each claim separately. A preference Elo summarizes human choices, not an objective accuracy rate.

At least two models must complete the same protocol and five identical tasks. We record version, settings and samples, and publish raw evidence plus the 0–100 conversion for every dimension before computing a score. Our category weights are editorial choices, not completed tests. Missing evidence means unranked; we do not turn provider specifications, arena Elo or individual API rates into invented scores. See the independent Artificial Analysis methodology pages for text, image, video and text-to-speech benchmarking; their results have not been reproduced here.

Independent reviews

A review should disclose task, prompt, model version, settings, sample count, cost and the limits of its observations. We separate subjective judgments from repeatable facts. No independent review has been published yet.

Corrections

When a source changes or becomes unavailable, we recheck the information and update the page date. Material corrections belong in the relevant article.

Independent methods: text, image, video, and text to speech.