OpenModels comparison
This version does not publish a composite intelligence score. It prioritizes source attribution, configuration separation, explicit missing data, and reversible history.
Canonical model YAML supplies names, descriptions, capabilities, legacy modality declarations, context limits, licenses, and registry-maintenance timestamps. Provider mapping YAML supplies deployment names, regions, context overrides and published prices. Registry timestamps are maintenance dates—not model release or measurement dates.
Unknown values remain unknown and never become zero. Context capacity is not quality. Provider facts stay attached to a deployment. A price row always names its provider and effective date. Derived developer or family identities are labelled until the registry supplies canonical fields.
The default price view chooses the active deployment with the lowest available input price, then output price, and displays that provider. A changed registry price closes the previous record with effectiveTo and appends a new version. Currency and optional cache, batch, reasoning, image, and audio fields remain separate.
Per-request cost is uncached input tokens times input rate, plus cached tokens times the cache-read rate (or normal input rate when absent), plus output tokens times output rate, divided by one million. Monthly cost multiplies this by request count. Tool, image, audio and unentered reasoning-token charges are excluded. Presets are editable workload assumptions, not benchmarks.
A result is admitted only with exact model and benchmark versions, score and unit, evaluation mode, reasoning effort, tool configuration, source, publication and verification dates, and methodology version. Official, independent, community and measured results remain distinguishable. Different configurations are never silently blended.
Comparable measurements must reference a model and provider deployment and record metric, unit, region, prompt type and version, sample size, distribution statistics, measured date, source, methodology version, and raw-result location. Current explorer views report Not measured until coverage meets this contract.
TTFT is request sent to first streamed token. Time to first answer token excludes hidden reasoning. Reasoning time, input processing, provider overhead, and end-to-end response are stored as distinct metrics. They are not interchangeable and are shown only for comparable prompt and deployment configurations.
Missing, not evaluated, not measured, partial, stale, and unavailable are separate states. Registry freshness is checked against a seven-day display policy; pricing should be verified at least weekly and when providers announce changes. A stale fact remains visible with a warning and its date rather than being fabricated or erased.
Corrections retain history and source evidence. OpenModels does not currently publish a composite intelligence score. One would require a public versioned methodology for inclusion, weights, normalization, exclusions and missing-data handling, with raw results preserved.
The public registry has incomplete canonical developer/family, modality-direction, release-date, lifecycle, commercial-use, structured-output and local-deployment coverage. Benchmarks and provider performance are not yet sufficiently populated. The explorer therefore launches with trustworthy registry facts, provider pricing and context—not synthetic quality claims.