Methodology

How VoidSource handles AI model data.

VoidSource combines public model data, public benchmark results, published API prices, selected first-party evaluation runs, and editorial judgment. The goal is practical comparison under real constraints, not a single magic score.

Model catalog

Model names, providers, variants, context windows, open-weight status, and capability notes come from public provider docs, API catalogs, model cards, and curated registry mappings.

Benchmark rows

Benchmark values are aggregated from public benchmark sources when available. Rows are only compared when the model identity, metric, variant, and evaluation conditions are compatible enough to sit in the same bucket.

First-party runs

When VoidSource reruns an evaluation, the page labels it as a first-party or reproduced run and keeps it separate from public leaderboard rows. These runs carry stronger evidence, but still include caveats.

Pricing

API pricing uses published list prices and generated pricing data. Self-hosted cost is not treated as a fixed provider price because it depends on hardware, utilization, power, runtime, and operational overhead.

How evidence strength works

Strongest

VoidSource first-party or reproduced runs, with source notes and caveats.

Useful

Public benchmark rows from known sources, when identity and conditions are comparable.

Context

Provider docs, model cards, pricing pages, release notes, and implementation notes.

Early signal

Community demand, comments, and news-derived evidence. These can guide research but should not pretend to be finished proof.

Editorial rules

  • A benchmark number is not automatically a recommendation. It is one signal inside a workload decision.
  • Different scaffolds, effort settings, API versions, or thinking budgets are not forced into one average.
  • Thin evidence is labelled as thin. Pages should say when there is not enough data to make a confident pick.
  • Editorial recommendations can disagree with raw score order when cost, latency, deployment friction, or risk changes the practical choice.
  • No paid placements. Commercial pages and service links are kept separate from ranking and evidence claims.