Tiers: how they are calculated

Every company and benchmark we track gets a tier, S to D, with its reasons, a confidence level and the steps behind it. This is exactly how a tier is decided; each list shows the tiers themselves.

S to D

Every company and benchmark gets one tier. S is a category leader or the standard benchmark; D is very early, or barely used yet. A tier always comes with its reasons, a confidence level and the steps that set it, on the entry's own page.

How we rate companies

  1. Adoption: team size and funding, from the curated directories and funding quoted from the company's own site or announcements.
    S
    Category leader: 200+ people, $100M+ raised or a $1B+ valuation.
    A
    Scaling: 51 to 250 people, or $20M+ raised.
    B
    Growing: 11 to 50 people, or $3M+ raised.
    C
    Early: a small team.
    D
    Very early: no disclosed funding or accelerator, and little presence on X.
    A company that went through an accelerator (Y Combinator and others) is never D.
  2. Curated directories. A company both RL directories list is at least C.
  3. Following on X. 25,000 or more followers on its own account moves it up one step, like a clear view of it does.

A company has one tier, shown in every section it sells into: RL environments, data, infrastructure.

How we rate benchmarks

  1. Adoption: who reports it, and how many models are scored on it.
    S
    The standard: frontier labs report it, or a trusted evaluator scores hundreds of models on it.
    A
    Widely used: a lab reports it, or 50 or more models are scored on it.
    B
    Established: 15 or more models scored, or curated RL directories rate it.
    C
    Niche: a handful of models scored, or built by an RL vendor.
    D
    Long tail: little or no usage data yet.
    Model counts from community leaderboards alone (thousands of fine-tunes) stop at A: S needs frontier labs or a trusted evaluator.
  2. Families. The root of a family (SWE-bench) sits one step below its best version (SWE-bench Verified): registries list usage under the versions.
  3. Curated directories. A benchmark the RL directories list together is at least B (two of them, at least C). They never lift a tier past that floor.
  4. Set apart. A leaderboard, arena or index that ranks models across many benchmarks is not itself a benchmark: it keeps its page but is not tiered.

For both

  1. What the market thinks, on X. Posts from the last month that name the entry are kept only when a model judges them really about it, then read as a whole for an overall view, how much buzz and a one-line take. A clear view on X moves a tier one step up or down; most entries don't move. The posts behind each read are linked on the entry's page.
  2. Confidence. How much stands behind a tier: usage, who backs it, a confirmed line from its own site or paper, and a clear read on X. High with three of the four, medium with two, low otherwise.

Where to see them

On every list, every company and benchmark page, and the map's Tiers filter: RL environments, data, infrastructure and benchmarks.

Rebuilt every morning; X is read again weekly. How every claim is checked.