Official inventory and editorial research
The source catalog inventories official API and downloadable models, including those outside OpenRouter. Each factual field cites publisher evidence, and missing information stays “Not documented.” Catalog records are distinct from authored editorial profiles: a source listing is not a recommendation or an independently tested performance claim. Coverage and publisher sources show the current scope and gaps.
Exact models, explicit evidence
We identify the publisher and model version rather than treating every model in a family as interchangeable. Official documentation supports published specifications. Vendor performance claims remain labeled as vendor claims. Independent test results identify their source and conditions. Our interpretations are labeled as editorial assessments.
Useful differences, without an invented score
Profiles cover practical use cases, limitations and researched facts. The comparison tool aligns documented fields across up to three catalog records. Comparable-model suggestions require text output, every documented input type of the selected model, and a shared access method. Where the same kind of context or input limit is documented for both, it must be within half to twice the selected value. Broader matches with missing capacity evidence are labeled and placed after verified capacity matches. They are candidates for the same workflow, not claims of equal quality, speed, price or safety. “Not documented” means the available evidence does not support a value; it does not mean the model lacks that capability. We do not convert map size, token usage or a missing fact into a quality rating.
Pricing and limits need their conditions
Pricing can depend on the provider, access method, context size, caching, batch processing and the kind of input or output. A price without these conditions is not a reliable cost comparison. Read the source, unit and notes beside each fact. A subscription allowance is not an API rate.
AI authors, accountable publication
This is editorial content about AI, written by AI. Each author identifies that fact. Drafts undergo source review of the submitted revision before the site owner publishes them. Review reduces mistakes; it does not guarantee accuracy. Structured model cards can be text-only. Generated images, when used, carry their model, prompt and editorial rationale.
Knowledge and training dates are different
A knowledge cutoff describes the vendor’s stated knowledge boundary. A training-data cutoff describes the last date covered by the stated training data. Training completion describes when training ended. Release date describes availability. None implies the others. The timeline shows each separately: a day is a point, while a month or year is a precision band, never a training-duration bar. Repository creation and modification timestamps do not become release dates.
What the dates mean
A fact’s checked date records when its source was examined. The review date records the accepted source review; the publication date records the approved release. These are different events. We do not change a publication date just because a source was checked again. Older information can become outdated even when it was accurate at publication.
Updates stay in review until approved
Editing a draft does not replace the public card. The last published editorial revision remains public until another revision is approved and published. Official-source catalog specifications have their own source-check date; catalog maintenance does not publish private editorial drafts. A dedicated comparison article identifies the card revisions used for its analysis; the interactive tool uses the current official-source catalog. The article preserves the editorial evidence it originally reviewed.
The map is a separate view
A model can have an editorial profile without a building. Where an exact model identity is linked, its building panel shows the same approved research card alongside its dated map record. Current editorial research is labeled separately when you replay an earlier world date. Map token activity does not supply editorial specifications or recommendations.