undominated-check

      ▄▌
    ▄██▌  Undominated.ai
  ▄████▌  undominated-check
▄██████▌  quote only

Is the model you are paying for beaten by something both better and cheaper?

npm install --save-dev undominated-check
npx undominated-check google/gemini-3.7-flash

One-shot, no install: npx undominated-check google/gemini-3.7-flash.

Read-only. It fetches published JSON from undominated.ai, prints the verdict, and exits. It sends nothing, stores nothing, and needs no key or account.

Not a router. It does not pick a model, hold keys, or execute inference.

Warn, never fail. Printing dominated still exits 0. Pass --exit-code only if a job must distinguish statuses — and even then, do not make it a merge blocker.

On a TTY the Pareto staircase and chartreuse paint the frontier finding only. Pipes, CI, --json, and --plain stay the same bytes as before — no colour in a log. --color forces the chrome when you are capturing a screenshot.

README badge

Same verdict as a static SVG. The slug is the dominance filename: / and : become __.

![Undominated](https://undominated.ai/badge/anthropic__claude-opus-5.svg)

Measured 2026-08-31: that URL returns HTTP 200, image/svg+xml. Nested /badge/anthropic/claude-opus-5.svg also 200; __ is the canonical prerendered key.

What it prints

$ npx undominated-check allenai/olmo-3-32b-think

  allenai/olmo-3-32b-think
  dominated — tencent/hy3 scores +142.6 and costs 39% less

  lens       lmarena
  workload   balanced
  as of      2026-08-30
  beaten by  https://undominated.ai/models/tencent__hy3/
  verdict    https://undominated.ai/data/dominance/allenai__olmo-3-32b-think.json
  page       https://undominated.ai/check/

  Something scores higher, costs less, and gives up nothing it can do.

Every line after the finding is there so you can disagree with it. The lens says which benchmark, the workload says which input:output mix priced it, as of says how stale it might be, and the verdict URL is the exact document this output was rendered from.

The five verdicts

| Status | Means | Exit code under --exit-code | |---|---|---| | on the frontier | Nothing in the catalogue is both higher scoring and cheaper. | 0 | | dominated | Something scores higher, costs less, and gives up nothing it can do. | 3 | | dominated, with a trade | Something scores higher and costs less, but drops context, output, a modality, tool use or reasoning. The loss is named. | 4 | | unrated | Nobody has published an independent quality score. | 5 | | unpriced | No published price, so no verdict is possible. | 6 |

unrated has its own exit code on purpose. Nobody having measured a model is not the same finding as a model being beaten. A gate that collapses the two will eventually approve a swap that no evidence supports. Absence of a verdict is never a favourable verdict.

dominated, with a trade is not a recommendation. Sixty-one percent of raw "better and cheaper" verdicts, measured on this catalogue, named a replacement that could not do the incumbent's job — half the context, or no image input. That is why the trade is a separate status with its own exit code, and why the losses are printed rather than summarised.

Usage

npx undominated-check <model-slug>
npx undominated-check --frontier

| Option | Effect | |---|---| | --local <dir> | Read from a directory instead of the network. Point it at a checkout's static/data. | | --origin <url> | Fetch from a different origin. Default https://undominated.ai. | | --json | Print the verdict document plus the resolved URLs, unformatted. | | --frontier | List every model nothing beats on both quality and price. | | --exit-code | Exit with the status-specific code above, for CI. | | --help, --version | |

Without --exit-code, exit is 0 whenever a verdict was printed, 1 on a usage error or an unpublished slug, and 2 when the data could not be read.

Offline

Nothing needs the network if you already have the files:

npx undominated-check --frontier --local static/data

  9 models on the frontier · lens lmarena · workload balanced · as of 2026-08-30

  upstage/solar-pro4                 1376.2  $0.0525/M
  qwen/qwen3-30b-a3b-instruct-2507   1384.3  $0.0844/M
  deepseek/deepseek-v4-flash         1431.6  $0.1013/M
  google/gemma-4-26b-a4b-it          1434.6  $0.1375/M
  tencent/hy3                        1441.2  $0.1444/M
  google/gemma-4-31b-it              1441.7  $0.1525/M
  xiaomi/mimo-v2.5-pro                 1465  $0.5437/M
  google/gemini-3.7-flash            1490.2  $0.75/M
  anthropic/claude-opus-5            1504.2  $10/M

  https://undominated.ai/frontier/

The directory needs frontier.json and dominance/<slug>.json, with / and : in the slug replaced by __. That is the same layout the site publishes, so mirroring it is a curl away.

In CI

Default is warn, never fail: a printed verdict exits 0, including dominated and unrated. A job that blocks a merge over a public-benchmark verdict under one workload is uninstalled the same day.

- run: npx undominated-check openai/gpt-5.2

--exit-code is opt-in, for jobs that want the status-specific codes above. Exit 3 or 4 means something cheaper now scores higher. Treat that as a prompt to look, not as a merge blocker — the verdict is computed on a public benchmark under one workload, and your workload is not that one. unrated (5) and unpriced (6) stay distinct from each other and from dominated: nobody having measured a model is not the same finding as a model being beaten.

What it does not do

• It does not compute anything. It reads the verdict undominated.ai already published. There is deliberately no second implementation of the frontier here, because two implementations eventually give two answers and only one of them gets checked.
• It does not price your workload. The published lens is balanced — three input tokens per output token, no cache. If your traffic is output-heavy or cache-heavy the ordering changes. /check/ lets you pick the workload; /audit/ reads it off an actual bill, in your browser.
• It does not know your constraints. Region, data-retention terms, latency, an existing contract and a rate limit are all invisible to it.
• It is not a router. It does not pick a model, hold keys, or execute inference. No affiliate links, no take-rate. See /independence/.

Data and licensing

The CLI is MIT. The data is not sublicensed by it.

Verdicts are computed over LMArena Elo, used under CC BY 4.0 from the official dataset — attribute LMArena if you republish a score. Each verdict document carries its own licence field saying so. No Artificial Analysis figure reaches this output; their redistribution terms are unresolved and the pipeline excludes them from every public surface.

Prices come from providers' own published rates via the OpenRouter models API.

Getting a number wrong

If a figure here disagrees with the vendor's own page, that is worth reporting and the correction gets logged rather than quietly overwritten: open an issue.

Licence

MIT. See LICENSE.