Model routers make educated guesses. ModelSpec makes informed, unbiased decisions from evidence.
ModelSpec is an analysis of alternatives for AI models: it decides which models fit a job from your requirements, sourced benchmarks and real cost, and shows its work. When one model wins, it says so. When the evidence can't separate them, it says that too, and returns the tie. Nobody pays to rank higher.
Same job. Tied on the evidence. $1,200 a month apart.
On coding alone, the evidence gives Claude Sonnet 5.5 a 39% chance of scoring at least as well as Claude Opus 5.5. At 25% or more, the evidence can't separate two models, so the board lists both as best for coding. At 10,000 tasks a month, Claude Opus 5.5 costs $2,400. Claude Sonnet 5.5 costs $1,200.
Published prices, 22 October 2026 snapshot. Your token counts change the numbers, and the board does the arithmetic in the open.
How ModelSpec compares to a router
Different job, different layer. Use both.
A router answers one question per request. ModelSpec answers the question before it: which models are worth routing to at all, and why.
| Compared on | A router | ModelSpec |
|---|---|---|
| The question | Which model for this request? | Which models belong on the list, and why? |
| When it runs | On each request, as it happens. | When you choose or review a model, or once per task for an agent. |
| Where it sits | In your request path. | Outside it. Your tokens go to the provider directly, and no prompts are kept. |
| What you get back | A call routed to one model. | A decision: the models that qualify, the ones screened out and why, how sure the evidence is, and what would change the answer. |
Decide what's worth routing to. Then let your router choose among those, request by request.
For teams and buyers
An analysis of alternatives, for every model choice.
When someone asks why you're on that model, the answer is a record, not a hunch: requirements, criteria, the alternatives and why each fell away, the evidence, its uncertainty and the cost. Each part links to how it works.
- RequirementsMust conditions are hard gates. A model that fails one is excluded, and stays on the board as excluded.Check it
- Weighted criteriaPrefer weights order the models that passed. They never remove one.Check it
- Alternatives screened, with reasonsEvery candidate is ranked, excluded with the reason, or listed as may qualify with the fact that's missing.Check it
- Measures of effectivenessAbility per domain, estimated from every admitted benchmark, with who measured each result.Check it
- UncertaintyA range on every estimate, and bands that say when the evidence can't separate the top models.Check it
- SensitivityTipping points, near misses, and which condition to relax when nothing qualifies.Check it
- CostCost per task at your token counts, plan break-even, and what you already pay for.Check it
- Audit trailA signed snapshot, a spec hash and a decision ID. Keep the spec, the snapshot and the decision contract version, and the same answer comes back next quarter.Check it
- IndependenceNo referral fees, no paid placement. The commitment is published, and checkable as data.Check it
What the board covers.
Decision snapshot 2026-10-22: the board covers 44 models (decider: 1, orderer: 2, text-generator: 35, vectoriser: 6). The frozen public catalogue has 1,372 cards, including retired models; 1,328 are not yet decidable. The Worker can serve a newer data image, so its health counts may differ.
Classes and offerings
| Class or placement | Decidable | Catalogued, not yet decidable |
|---|---|---|
| actor | 0 | 0 |
| analyser | 0 | 48 |
| decider | 1 | 0 |
| derived | 0 | 1 |
| forecaster | 0 | 12 |
| labeller | 0 | 19 |
| media-generator | 0 | 75 |
| orderer | 2 | 11 |
| scorer | 0 | 8 |
| simulator | 0 | 0 |
| text-generator | 35 | 979 |
| transcriber | 0 | 21 |
| unclassified | 0 | 5 |
| vectoriser | 6 | 149 |
Covered API offerings: 43 across Alibaba Cloud Model Studio (1), Anthropic API (7), Amazon Bedrock (3), Azure AI Foundry (4), DeepSeek Open Platform (2), Gemini API (Google AI Studio) (5), Vertex AI (Google Cloud) (11), Meta Model API (2), OpenAI API (4), TypeSafe API (1), xAI API (1), Z.ai API (2). Router/aggregator offerings in this snapshot: 0 (providers: none).
Generated coverage, keyless. Catalogue presence does not establish decision coverage.
The right model for every role in your agent stack.
One call per role picks the model that fits it, explains why, and gives the same answer every time for the same facts.
Three ways in: keyed CLI, MCP and HTTP API. Machine data access needs a ModelSpec API key; only human lookup is free. No data download or local decision cache.
uvx --from modelspec-dev modelspec pipx install modelspec-dev pip install modelspec-dev
pip install modelspec is an unrelated project. Use modelspec-dev.
Read the agent guideMachine access requires an API key.
Think you know the best coding model?
Every number is one click from its source.
Which benchmark, which date, who ran it. Independent results sit beside the lab's own claims, and each is labelled.
How we decideUnknown means unknown.
A model with no published answer to your question stays on the board as "may qualify". It never becomes a zero, and it never quietly disappears.
How unknowns workNobody pays to rank higher.
No referral fees, no paid placement, no sponsored slots. It's a published commitment you can check.
Read the commitmentCatalogue and machine access
ModelSpec catalogue and decision snapshot
The frozen public image of ModelSpec model cards and benchmark evidence. Current model selection answers are computed per request by the API.
HTTP endpoints to decide, rank and check licence and residency policy.