AI station rankings and model-use guide

Find the model that works best.

Choose a model, then compare the station routes that serve it. Availability, P95 response time, price, evidence, and assessment window stay attached to each route.

View the current-model rankingsOpen Quick Test
Current model / Route preview—Reading the public snapshot…
Comparison framework

01 / Availability

Measure a window, not a moment.

Availability stays tied to the model, route, sample, and observation window. One successful request is not uptime.

  • Versioned public snapshots
  • Evidence labeled by source
  • Paid placement cannot change rank

Only conclusions supported by the current snapshot are shown. Missing and unverified fields stay visible.

Model ecosystem / Current catalog

Recognize the model family before comparing the route.

These marks identify model families present in the current public model catalog. They are not partnership badges and do not affect rankings.

model entries in the public catalog
—
identified model families
—

Reading the public snapshot…

Product names and trademarks belong to their owners. Every mark receives the same size and fixed family order. A catalog entry does not by itself mean measured coverage, endorsement, or participation in Folkbench evaluation.

Comparison framework

One framework. Six different questions.

Folkbench separates operational facts, prices, provenance, and evidence limits so one polished score cannot hide what is unknown.

The illustrations explain field logic; they are not published measurements. Model, station, channel, method, sample, and time window remain one comparison context. Change one and the conclusion must be reconsidered.

01 / Availability01

Measure a window, not a moment.

Availability stays tied to the model, route, sample, and observation window. One successful request is not uptime.

02 / Latency02

Read distributions, not a best case.

P95 full-response latency exposes the slower edge of completed requests; missing samples remain missing.

03 / Consistency03

Treat identity as a bounded signal.

Response consistency can flag review candidates. It does not prove which model produced a response.

04 / Price04

Compare like units only.

Currency, input/output unit, route, and update time must align before prices are compared.

05 / Provenance05

Name who supplied the fact.

Platform measurement, vendor fact, public record, and user report stay visibly separate.

06 / Missingness06

Do not manufacture precision.

Expired, insufficient, unverified, or incomparable fields are shown as such instead of being estimated.

Current snapshot / Top five

The top five are a shortlist.

Overall score narrows the field. Availability, P95 latency, price, and evidence determine whether a route fits your task.

View all rankings
Current model / Route preview—

Loading the current published preview…

Quick Test / Controlled entry

Quick Test answers one attempt.

It checks one endpoint in one limited request. It does not create a public rank, long-term SLA, or security certification.

Current boundaryBeta preparation flow
  1. 01

    Choose context

    Only public route context moves to the test page. Endpoint and key stay out of the URL.

  2. 02

    Confirm authority

    Sign in, confirm permission to test, and review the data notice.

  3. 03

    Read one result

    The result describes that attempt only and never enters the public ranking.

Real probe execution is not enabled. This entry shows preparation only and stays unavailable when required services are missing.

Open Quick Test preparation

Evidence pipeline

Every public number must be traceable.

A result is meaningful only when object, source, sample, window, method, and limits remain attached.

Snapshot
Not provided
Method
Not provided
Window
Not provided
Updated
Not provided
  1. 01 / Source

    Name the source

    Keep platform measurements, vendor facts, user reports, public records, and unverified claims separate.

  2. 02 / Window

    Attach time and sample

    Record route, model, configuration, sample range, and observation time.

  3. 03 / Method

    Keep gaps visible

    Version the method and expose missing, expired, insufficient, or incomparable fields.

  4. 04 / Publish

    Freeze the reviewed view

    Publish provenance and limits with the fact; keep corrections auditable.

Evidence ceilingAn anomaly shows what happened in one observation, not why. It cannot replace a long-term SLA, audit, or certification.

Independence rules

Visual presence never becomes score weight.

Authority comes from published rules and traceable facts, not from a logo wall or a commercial relationship.

Visible elementPublication rule
01Model logosIdentify catalog families only; they imply no affiliation, endorsement, evaluation participation, or rank weight.
02Vendor factsStay labeled as vendor supplied until a separate source and review support a stronger status.
03Commercial relationshipsRemain separate from scores, ordering, evidence status, and correction decisions.
04Missing dataRemain missing; facts from another model, channel, or station are never used as a substitute.
05Quick TestDescribes one authorized attempt and never enters a public ranking release.

The active published snapshot is the landing page's only ranking source. Drafts and commercial placement stay outside that read path.

Continue with context

Choose the next step by question.

Compare in Rankings, understand the method in Guide, or correct vendor facts.

Guide

Read the ranking

Understand each field, board, and evidence limit before comparing.

Open the guideVendor

Correct vendor facts

Request an addition or correction; submissions never change rank automatically.

Open vendor onboarding

Start with one model

Compare first. Verify next.

Use the snapshot to find candidates, then inspect each route before deciding.

View rankingsOpen Quick Test