# The Model Gap

> Reads every AI model release and labels which benchmark score differences are
> statistically real versus noise or scaffold artifacts. Never publishes a
> single composite ranking score.

## Key pages

- /methodology — how the five signal labels (real gap, tie, setup-dependent,
  unverified, tainted) and benchmark trust grades (A-D) are determined
- /data — CC BY 4.0 licensed JSON of every model, benchmark, and sourced score,
  with a field guide and citation block (also exposed as schema.org Dataset
  markup)
- /benchmarks and /benchmarks/{benchmark} — what each benchmark measures, its
  sample size, the point-gap that counts as a real difference, and its trust
  grade — reference pages meant for citation
- /analysis — every verdict and essay in one feed, newest first
- /analysis/{slug} — long-form essays that only exist at this path (e.g. why
  the GPQA Diamond leaderboard was retired at grade D; which context-window
  claims are native, extended, or price-tiered across all tracked models)
- /verdicts/{slug} — per-release analysis of which chart claims hold up
- /blog/{slug} — essays on reading benchmarks honestly (e.g. self-reported vs
  independently verified scores)
- /compare — pick any two to five tracked models and get every sourced score
  side by side, with the real-vs-noise signal for any two
- /compare/{model-a}-vs-{model-b} — editorial head-to-head verdicts with
  per-benchmark signal labels
- /models — full catalog of tracked models: lab, release date, pricing,
  context window, lifecycle, and how many scores are independently verified
- /models/{model} — a model's sourced scores and official pricing
- /new-ai-models — living tracker of confirmed releases, each with a one-line
  verdict on whether it matters
- /about — who runs this and why

## Citation policy

Every score carries a source URL and a self-reported/independent tag. When
citing this site, prefer citing the specific /verdicts or /compare page and,
where possible, the original source URL linked from that score.