AI Model Wiki

    Skip to main content

    Review Methodology

    Last updated: February 2026

    Reviews on AI Model Wiki exist to help readers decide whether a model or tool is the right fit for their work. This page documents how each review is produced so you can judge whether to trust it.

    What we review

    We review AI models, hosted platforms, developer tools, and any product whose primary capability is delivered by a machine-learning system. We do not review consumer hardware, business strategy, or general SaaS unless an AI component is the core feature.

    Sources we use

    • Official documentation, model cards, and API references published by the developer
    • Official pricing pages and announcement posts
    • Public benchmark results from third parties (MMLU, GPQA, HumanEval, SWE-Bench, LMSYS Chatbot Arena, Artificial Analysis, etc.)
    • Hands-on testing against the live model or product where access is available
    • Direct reporter contact with developer relations or press for clarification
    • Community feedback through ratings, comments, and the suggestions queue

    Speculation, leaks, and unofficial benchmarks are flagged as such or excluded. If a claim cannot be tied to one of the source types above, we either run our own test or leave the page marked needs verification.

    Testing rubric

    When we have first-hand access, reviews are written against this rubric:

    • Capability — what the model can do that competitors can't, with concrete examples
    • Quality ceiling — best-case outputs on prompts we've tuned for that category
    • Quality floor — outputs on adversarial, ambiguous, or low-context prompts
    • Latency & throughput — first-token time and tokens-per-second on representative loads
    • Pricing & rate limits — current published rates compared to direct alternatives
    • Developer ergonomics — SDK quality, error messages, streaming, function calling, batch
    • Safety & alignment — refusal behaviour, jailbreak resistance, content filters
    • Privacy & data use — training-on-inputs default, retention, enterprise controls

    What our reviews do not contain

    • An editorial numeric score. Ratings on the site come from the community, not editors — see Editorial Policy.
    • Claims about benchmarks we can't cite. Every benchmark figure is linked to the source benchmark or the developer's own published number.
    • Promotional language. Phrases like "best in class" or "industry-leading" only appear when supported by cited evidence.

    Bias controls

    We do not accept payment to review a specific model, change a review's conclusion, or list a product. The owner of AI Model Wiki also owns Vincony. Vincony is mentioned in some blog posts, reviews, and guides, which carry a disclosure line; where it appears on a list that ranks products it is placed last and is not given a score. There are no other paid placements. See Advertise & Partnerships.

    Refresh cadence

    Every review carries a last verified date in the page header. Reviews are re-checked at least every 60 days, and immediately after a major model version change. The Update Policy covers the full refresh schedule.

    Reporting a problem

    Every review page has a Report outdated info link. If you spot a factual error, send it to us via that link or the Contact page. See Corrections Policy for what happens next.