LPH98/undominated-ai-model-pricing

Undominated AI model pricing and Arena scores

Snapshot of 442 model listings from 53 providers, measured 2026-10-13. Every row carries its own source URL and fetch date.

LLM catalogue only: image, video and embedding catalogues are not part of this dump.

Produced by https://undominated.ai · scripts/build-export.mjs

What is in here

Rows	442
Providers	53
Rows with a price	433
Rows priced at zero	16
Rows with context-tiered pricing	70
Rows with time-of-day pricing	2
Open-weight rows	33
Rows with an announced retirement date	19
Rows with an LMArena Elo	144
…of those, with a confidence interval	144
Rows cross-checked against a vendor pricing page	46

What is not in here, and why

298 of 442 rows carry no quality score in this deposit — 67% of the file. That absence has two different causes and they are not interchangeable:

• 150 rows are unmeasured. Nobody has scored them.
• 148 rows were measured and the score is withheld here. It comes from Artificial Analysis, which this deposit may not redistribute. The model is not unmeasured; this file simply may not carry the number.

In neither case does null mean zero, and sorting either group to the bottom of a ranking would defame a model on the strength of our licence rather than its capability. Read null as "not stated here".

396 of 442 rows have no vendor cross-check. Their price came from OpenRouter's listings and nothing else. provenance.enrichedBy is the vendor document URL where a human checked the model-level listing (modelLevelPricing, or pricing on a model-level row), and null where none did. Do not read the presence of a price as confirmation of it.

No Artificial Analysis field is present. AA's free tier permits internal use only; public redistribution requires a commercial licence this project does not hold. Their intelligence, coding and agentic indices reach our internal dataset because OpenRouter re-serves them, and passing through an aggregator does not change the terms. The fence is an allowlist in scripts/build-export.mjs, so a new AA field upstream is excluded by default rather than by remembering to exclude it.

No model descriptions. Prose is somebody's copyright in a way a price is not.

Two upstream provider entries are withheld from every published surface: one is a test fixture in the upstream registry, the other serves a zero-priced model with no privacy policy. Neither is a real seller anyone can buy from.

Files

File	What it is
models.json	Every exported field, nested, one object per listing.
models.csv	The 34 headline columns, flat.
datapackage.json	Frictionless-style manifest: schema, sources, and a SHA-256 per file.
LICENSE.md	Terms and required attribution.

How it was measured

pricing is the reference price: of the endpoint rows OpenRouter lists for a model, the cheapest on the balanced blend (3 input tokens to 1 output) that is serving, at standard delivery, is priced at its standard time window, is at the maker's precision or better (a 4-bit row counts only when the maker itself serves 4-bit, and a row that declares no precision counts only when it is the maker's own or the model's weights are closed), and serves at least 90% of the reference context: the context the maker's own serving rows serve, the smallest where they differ, or, where the maker does not serve the model, the context length published for it (contextWindow). A row is serving when OpenRouter's status for it is 0 or above, or reads degraded (−2) while the row answered at least 95% of the last day; and never when it answered under 80% of the last day, whatever it reads now. A row on promotion is counted, and published, at its standard rate: each rate ÷ (1 − discount), with the fraction in priceRow.discount. The row's whole rate card is published, never rates mixed from two rows. priceRow names the row; priceBasis is reference, deal-only (no row passed: a fallback, the cheapest serving row of standard delivery from a reseller that declares no precision, failing that a 4-bit one, each at its standard rate, with why it is not a reference price in priceRow.reasons) or model-level (no row to take a price from, or a price held at the last published record while a newer upstream value is reviewed: pricing is OpenRouter's model-level listing as last accepted). Where the price is taken from a row, modelLevelPricing keeps that model-level listing beside it; a model-level record has no such field, because pricing already is it. deal is the cheapest excluded row that is cheaper right now, with why it was excluded, or, where the price row is itself on promotion, that row's promoted price; it never sets a rank. A reason is promotion (with pct), delivery-tier (with tier), four-bit (with quantization), precision-not-disclosed: a reseller that declares no serving precision, which is not full precision and is not known to be less; or context-below-reference (with context, the tokens the row serves): under 90% of the reference context. Rates are normalised to USD per million tokens. Where a seller charges different rates above a context threshold, every rung of the ladder is kept in pricing.tiers — flattening a ladder to one rung mispriced five models by up to 2× once, which is why the rungs are all here. Where a vendor varies price by hour, pricing.priceWindows carries each window with its own rates rather than a single averaged figure.

LMArena Elo is the highest-scoring Bradley-Terry row from the text arena, with the reasoning effort that produced it in benchmarks.lmarenaEffort. A score without its effort level is not comparable to another score; a model's rank depends on how hard it was allowed to think. benchmarks.lmarenaCi carries Arena's own confidence bounds, so two scores whose intervals overlap should be read as a tie, not as a ranking.

How to cite

Undominated.ai (2026-10-13). Undominated AI model pricing and Arena scores. DOI assigned on deposit — see datapackage.json for the content hash that identifies this exact snapshot.

The manifest in datapackage.json carries a SHA-256 for every file and one combined contentHash over all of them. Cite that hash and a reader can confirm they are looking at the same bytes you were.

Licence

Arena-derived fields are CC BY 4.0 and carry a required attribution. Everything else is set out in LICENSE.md. Read it before redistributing.

Corrections

This project publishes the claims it got wrong, dated, rather than quietly editing them. The log is at https://undominated.ai/corrections/ and in docs/EMPIRICAL-GROUNDING.md in the source repository. If you find an error in this snapshot, it belongs in that log — say which row and which field.