Urdais Open-weight vs Proprietary Methodology

Version 1.0.1, approved 16 September 2026. Status: approved for production, effective from 16 September 2026. Supersedes 1.0.0, which remains on record. Prepared from the Phase 4A codebase and production-substrate audit, and approved after the implementation was built and verified against published publisher artifacts.

1.0.1 is a wording-only clarification of §8: it splits the explanation of unlinked volume into identifiers Urdais has not mapped and identifiers that cannot be resolved at all. It changes no number. The public classes, the denominator, the fold, the capability rule and the price rule are byte-identical to 1.0.0. A successor version exists rather than an edit in place because an approved version's content hash pins the document as it was approved, and re-pointing it at a later edit would destroy the history the hash exists to keep.

Open-weight vs Proprietary has no ingestion of its own. It is a classification laid over data three other Urdais products already collect: observed token volume from UTVI, benchmark capability from Epoch AI, and list price from Urdais Token Price. The only thing this product persists is the classification itself, because a licence finding is evidence rather than a derived statistic.

1. The claim

Whether each model's publisher released downloadable weights, measured across observed OpenRouter token volume, Epoch AI benchmark scores, and published list prices.

That is the whole of it.

2. The semantic boundary

Three over-readings are refused explicitly, because a reader would otherwise make all three for free.

Open-weight means the publisher offers the weights for download under a licence permitting commercial use. It does not mean open source, and several of the licences counted as open-weight here restrict commercial use in some conditions. Weights published for non-commercial or research use only are reported as Unclassified, not as Open-weight.

Volume is share of observed OpenRouter token volume, not of the industry.

Prices are published list prices per token, not total inference cost: a model that emits more reasoning tokens to answer the same question costs more than its per-token price suggests.

Urdais does not claim, on any surface, in any label, tooltip, heading or export:

that open-weight models are better or worse than proprietary ones
that an open-weight share of observed volume is an open-weight share of the AI industry
that any model in either class is lawful for a given reader to use
total inference cost, cost per task, or the cheapest model to accomplish an outcome

The third deserves a sentence of its own. A licence name is a citable fact and is published; whether a particular licence permits a particular use is legal advice about a specific reader, and Urdais does not give it. The fourth is inherited from Model Frontier, where the same list-price arithmetic is bounded the same way and for the same reason.

3. The classified unit

A classification attaches to one exact canonical model version, never to a lab, a brand, a family or a name.

Production already holds the counterexample. One publisher ships models under a single brand and a single version number where some are downloadable and some are not, and a schema keyed on anything coarser would make that distinction unrepresentable. A second counterexample appeared during this phase: the same model name carried a restrictive community licence in its April preview and a permissive licence at its July full release. Classifying "that model" would have been wrong in one of the two directions whichever answer was chosen.

4. The taxonomy

Six internal classes are recorded, because the licences are that different:

open_weights_unrestricted — weights published under a licence with no field-of-use restriction.
open_weights_restricted — weights published under a licence that restricts some commercial use, by revenue threshold, user count, attribution obligation or acceptable-use clause.
open_weights_noncommercial — weights published for non-commercial use only.
api_only_closed_weights — offered as hosted access; the publisher lists no downloadable weights.
unknown — Urdais has not established how this model's weights are distributed.
not_applicable — the source does not disclose which model served this volume.

Three public classes are shown:

Open-weight — open_weights_unrestricted and open_weights_restricted.
Proprietary — api_only_closed_weights.
Unclassified — open_weights_noncommercial, unknown and not_applicable, plus volume that never reached a classification at all, plus the source's own residual row.

The fold is stated rather than hidden. Restricted weights roll up to Open-weight because the obligations they impose — a notice above a revenue line, an attribution in a product's interface — still leave a model that can be self-hosted, fine-tuned and sold from. That is what changes the economics this section measures.

Non-commercial weights do not. They are downloadable, which is why they are an open_weights_* class internally, but they may not lawfully serve the paid inference whose volume and price this section compares. Counting them as Open-weight would place them inside a commercial comparison they are excluded from, and would inflate the open-weight share with traffic that cannot legitimately exist. They are reported as Unclassified, counted separately from the other unclassified causes, and the internal class is preserved so the distinction stays recoverable.

The source's own residual row stays separately identified throughout. It is the source-defined residual — volume OpenRouter aggregates without naming a model — and it is never merged with the identities Urdais failed to resolve, because one is a property of the source and the other is work Urdais has not done.

5. Evidence rules

A classification requires a positive finding about a publisher artifact. The hierarchy, strongest first: the weight licence itself, the publisher's own weight repository, the publisher's model card, release documentation, an official repository, and an official pricing or access page.

Failing to find weights is not evidence that none exist. The schema enforces this rather than trusting discipline: an open class requires a locatable artifact URL and one of the stronger evidence types, and the evidence type no_evidence_found may only ever support unknown.

Proprietary is a positive finding too. It means the publisher enumerates how the model is obtained and every channel is hosted — not that a search came back empty. Where a publisher's own page lists five hosted platforms and offers no download, that enumeration is the evidence.

Where corroboration is weaker than a direct publisher statement, the classification row records that in its own column rather than presenting the finding as stronger than it is. One model in the initial set is recorded as unknown for exactly this reason: secondary sources describe it as hosted-only, no first-party statement was located, and suggestive is not positive.

6. Classification changes, and the old answer was not wrong

A model released hosted-only may have its weights published a year later. So a classification is superseded, never updated in place, and carries the date from which it is true — which is what lets a historical statement stay defensible after the world moves. The read layer joins only live rows.

7. Identity

Volume arrives from UTVI as the source's permaslug, verbatim. Connecting it to a canonical model is an evidenced link or it does not happen. There is no fuzzy matching, no alias guessing, and no fallback that assigns volume to a model on resemblance.

Two link rules are used, and each link records which one applied:

Where the publisher versions by date — DeepSeek names its own checkpoints 0423, 0731, 0813 — the date is part of the identity and a mismatch is a real mismatch.
Where the publisher's identifier is dateless and the name is unique, the source's date suffix is a listing date and the link is by name.

Unlinked volume keeps its tokens and loses its class: it is reported as Unclassified, never assigned to either side.

Share of observed token volume over a trailing 30-day window, anchored to the newest published UTVI date rather than to today, so a day when collection has not yet run shifts nothing.

The denominator is total observed tokens, the same one Market Share 1.1.0 uses. Dividing by classified-only volume would make the two known classes sum to 100% while a large share of traffic sat outside the calculation, which turns "of what we could classify" into "of the market" in the reader's head. Two Urdais products disagreeing about the size of the market would be worse than either answer.

Unclassified is a published number, not a rounding error, and its four causes are reported separately because they are different work:

source-aggregated — the source-defined residual: the source's own aggregate tail row, which names no model and can never be resolved by Urdais.
unmapped — a source permaslug that exists but carries no evidenced canonical link yet. Genuine research debt, and the number that should fall as identity work is done.
unresolvable identity — an identifier that cannot be defensibly resolved to the exact model that served the observation: an alias, an anonymous or stealth endpoint, a routing identity. Not unfinished work. No amount of research produces a model here, because the source names none.
undetermined access — the canonical model is known and its access class is not yet established. Evidence work.
non-commercial — a linked model whose weights are published for non-commercial use only. A settled finding rather than missing work, and the one Unclassified cause that looks like Open-weight from the outside.

Unlinked volume is therefore split between identifiers Urdais has not yet mapped and identifiers that cannot be defensibly resolved to an exact canonical model, such as anonymous or routing aliases. Reporting the two as one figure would imply that all unresolved identity is a backlog that effort will clear, and a large part of it is not.

The largest single unresolved identifier in the initial set is an anonymised stealth endpoint. Its publisher later stated publicly which model had been behind it, but the permaslug names a routing alias rather than a published model, and Urdais cannot establish which model served the volume recorded under it on any given date. Its volume is kept and reported as unresolvable identity rather than assigned to a class on a press statement, and rather than counted as research Urdais has yet to do. A later public attribution does not convert an alias into an evidenced link.

9. Capability gap

For each benchmark, the highest-scoring configuration in each public class, and the difference between them.

The unit is the same one Model Frontier plots: a canonical priced model SKU under one source-declared configuration. Ties resolve on the canonical model identifier, so the reported model does not depend on row order. Only models carrying a live classification in one of the two public classes participate; a model that folds to Unclassified has nothing to contribute to "the most capable open-weight model," and the production check reports how many configurations that excludes on each benchmark.

Where one class has no measured model on a benchmark, the gap is not reported. There is no comparison to make, and a gap against an absent side would be an artefact of coverage.

10. Price gap

Median blended list price per 1M tokens in each class, taken over the Pareto-efficient configurations of the Model Frontier for the selected benchmark.

The population is not selected here. It is exactly the set of configurations Model Frontier marks efficient — the same objects, from the same derivation, through the same markFrontier call the chart uses. There is no capability threshold, fixed or derived. A second selection rule in this section could drift from the one the chart draws, and then two panels of one page would disagree about which models are efficient.

Configurations remain the unit, and are never collapsed first. Domination is configuration-level: a model can be efficient at high reasoning effort and dominated at low effort. Deduplicating to one price per model before the efficient set is determined would decide which of a model's configurations speaks for it — exactly the choice §9 and the Model Frontier methodology refuse to make.

Three further rules:

The sample is reported. Each class publishes how many efficient configurations it contributed, as n.
Each class is reported independently. A class with no efficient configuration shows — and n = 0; its absence never suppresses the other class's median. One-sided coverage is itself a fact about the frontier, and withholding the side that does exist would publish less than the data supports. No fallback median is substituted for the absent side.
A headline ratio is published only when both classes carry at least three. Below that, the section renders Insufficient comparable frontier coverage and publishes the per-class medians and samples without a multiple. A median over one or two points is an artefact of which models happened to be evaluated and priced, and a ratio drawn from it would be unfalsifiable in a different way from a tunable threshold but no less unfalsifiable. Thin coverage is an expected product state; widening the population until a ratio appears is not an available response.
The median averages the two middle values when the count is even. Either convention is defensible; this one is stated so the number is reproducible from the published rows.

The price of each class's highest-scoring configuration is carried beside the median as a clearly labelled companion metric. It is one configuration's price, not a central tendency, and it is never the headline.

The price is the same 50/50 input-output blend Model Frontier plots, computed by the same function, so a price quoted in one section cannot disagree with the same model's price in the other.

11. What is derived at read time

Everything except the classification. Volume shares, capability bests and price medians are aggregations of rows the platform already holds, recomputed on each request from live publications, live snapshots and live classifications.

This is how revision is handled — structurally, rather than by procedure. A revised UTVI date, a superseded capability observation or a superseded classification changes what this section returns on its next read, with nothing to invalidate and nothing that can go stale.

12. Refusal states

The section publishes nothing rather than something wrong. An unconfigured deployment, an unreachable database, a methodology version still in draft, or a trailing window with no observations all produce an explicit statement that no comparison is published.

A benchmark where one class has no classified model is reported as a benchmark whose comparison cannot be made. That is a real and reportable state, not a broken section.