V2 live Sep 6, 2026 5 peer boards · 5 evidence outputs · legacy scores isolated 3 changes in this edition →
EDITION 2026.09-V2 / EVIDENCE ARCHITECTURE
A visual builder, an IDE, an autonomous coding agent, a desktop agent platform, and a model are not one contest. V2 gives each a legitimate peer board and keeps quality, capability, user outcomes, adoption, price, coverage, and uncertainty separate.
No universal winner. Missing evidence stays unknown, adoption never becomes quality, and close results publish as rank bands. Read Method V2 →
THE NEW CORE / FIVE PEER BOARDS
Products may appear in more than one board, but every placement has its own eligibility rules, evidence basket, coverage, and uncertainty.
Loading Method V2…
Twelve seconds. Then everything else on the page makes sense.
5 QUESTIONS / NO ACCOUNT
Answer five plain questions. The finder transparently re-weights the published evidence model and returns exactly three options, each with a reason and a tradeoff.
Pick an answer to the first question to see recommendations.
EVIDENCE LEDGER / TYPES NEVER SILENTLY BLENDED
Every observation names its type, source, scope, age, and confidence. Discovery order is not a quality rank, and raw browser votes have no ranking authority.
Discovery order is not an AI Scoreboard quality rank. It uses visible catalogue facts to expose records and gaps. Historic browser reports are retained only as an excluded pilot audit trail; every card states what is known and what remains unverified.
LEGACY 2026.09 / EDITORIAL LENSES
These reweighted editorial subscores are preserved for transparency, but they are no longer the V2 universal ranking.
Balances visual quality, engineering depth, speed, control, ecosystem, delivery, and value.
USER MOMENTUM LEADER
49.665 trillion tokens and 500.5 million requests. Hermes is also #2 on the day, week, and coding-category views. That is extraordinary verified usage, not a claim that it has the best visual output.
Trailing windows, not downloads, stars, or survey opinions.
Products not routed through OpenRouter are marked not measured, never zero.
Future ratings will shrink small samples toward the catalogue mean and decay stale activity.
Public ranks can contain gaps because hidden and private apps are excluded. We publish the source rank as-is and never densify it. OpenRouter apps ↗ Ranking definition ↗ Bayesian method ↗
LEGACY SNAPSHOT / NOT METHOD V2
20 ranked tools shown
| Rank | Tool | Score | Dimension profile | Why it ranks here | Price / free tier | Code export | Best for | Conf. | Compare |
|---|
These entries are in the index and searchable, but the rubric evidence needed for an ordinal rank does not exist yet. No score is invented, and published ranks are untouched.
Ranked entries are commercial or open products with first-party frontend evidence. Terminal-only agents, spec-driven PR bots, and brand-new releases enter as tracked entries first. Nominate a missing tool by opening an issue against the published method.
INTERACTIVE MODEL
Change what matters and the board re-ranks instantly. The model stays constrained to 100%, and the preset is written into the URL so you can share the exact view you are looking at.
Overall = Σ (subscore × dimension weight ÷ 100)
Whole-number subscores; overall rounded to one decimal. Other sliders normalise proportionally when one changes. Published ranks always use the Balanced preset.
CATEGORY LENS
Different jobs, different winners. Category context prevents false equivalence.
LEGACY MODEL LENSES / EXPLORATORY ONLY
These controls explore the archived composite. They do not create Method AIS-V2 ranks; current model claims require versioned model × harness × settings × task records.
QUALITY / COST FRONTIER
Up is better. Left is cheaper. The models on the glowing edge deliver quality that cheaper rivals cannot match within the evidence tolerance.
Calculating the current quality premium…
MODEL EVIDENCE DIRECTORY / 100 PROFILES
Loading 100 models…
These records match the same search above. They have verified availability metadata, but no ordinal quality rank until benchmark evidence is normalized across comparable harnesses.
CATALOGUE / NOT A COMMON RANKING
Tools and models are grouped by type. Each score is shown only inside its own evidence model.
A product workflow and an underlying foundation model are different units. Combining them would create false precision, so the catalogue keeps them apart and labels the type on every card.
DECISION WORKSPACE / PEER-AWARE
Select a need and legitimate peer group first. V2 compares evidence families side by side and says best fit, never objectively best.
WHAT CHANGED / EDITION 2026.09
Every edition states what moved and why. Subscribe with the Atom feed.
EDITION HISTORY
30 tools and 100 models. Entries are labelled New, Updated evidence, Price changed, Tracked, or No prior snapshot. Rank deltas begin only after a comparable numeric snapshot exists.
No rank movement is claimed for this edition.
METHOD AIS-V2 / SOURCE QUALITY
V2 separates products, harnesses, models and evidence records. It normalizes only comparable runs and publishes coverage, uncertainty, and rank bands.
quality-weighted mean of same-version peer percentiles + shrinkage
Sparse evidence moves toward the peer mean. Missing evidence is unknown, never zero.
weighted geometric composite + published coverage ratio
Critical weakness cannot disappear behind one exceptional benchmark. Cost uses workload scenarios and a strict Pareto frontier.
Official product, pricing, and documentation pages first; primary financial disclosures second; credible reporting and independent hands-on evaluations for corroboration and limitations.
Ranked requires two independent evidence families, two pillars, at least 60% coverage, one moderate-or-higher record, and a qualifying observation no older than 180 days.
Missing facts stay n.a.; conflicts stay visible. Corrections are welcome and are logged in the next edition's Updates section.
EVIDENCE CARD
SIDE-BY-SIDE
COMPARISON ALREADY OPEN