Research methodology

How we turn model claims into useful comparisons.

1. Source hierarchy

We prioritize benchmark-maintainer results and reproducible technical reports, then official model cards and provider documentation. Third-party aggregators can help locate a claim but are not treated as the primary source. Each model page links to a first-party reference.

2. Version control

A model name is attached to a specific release where possible. Preview aliases and silently updated endpoints are flagged during review. We record the page review date because API behavior and price can change without a model rename.

3. Benchmark handling

Scores remain separated by benchmark family. We preserve missing values and use reported percentages without extrapolation. Prompting, tools, sampling, and harness versions can prevent direct comparison; readers should inspect the originating report for experimental settings.

4. Pricing

Prices are shown per one million text tokens in US dollars when a stable public API price exists. Batch discounts, cache reads, reasoning-token accounting, regional taxes, and hosting costs may change the effective amount.

5. Open-model language

“Open weights” means model weights can be downloaded under stated terms. It does not imply an OSI-approved license, unrestricted commercial use, released training data, or reproducible training.

6. Selection and update cadence

We add models with meaningful adoption, novel capability, deployment relevance, or strong evidence. The directory is structured for a weekly source review; each published snapshot must reflect the date actually checked.

7. Corrections

Send a source URL and description to info@isrange.com. We verify against primary material and update factual errors promptly. Opinions may be revised when better evidence becomes available.

Methodology version 1.0 · Effective July 19, 2026