Model Catalog

Intutic ships a generated catalog of LLM models — which provider offers each one, its context window, output token limit, per-1K token cost, and whether it's a reasonable fit for judging another model's work. It's available from @intutic/shared-types for anything you build against Intutic's provider registry.

Where it comes from

The catalog is generated from the same upstream source that prices every request this proxy handles — LiteLLM's model_prices_and_context_window.json — via tools/scripts/build-offline-pricing-bundle.ts. One fetch produces two files:

• packages/proxy/src/pricing/offline_bundle.json — compiled into the proxy binary, used for cost estimation.
• packages/shared-types/src/modelCatalog.generated.ts — the catalog this page documents.

Both are regenerated together, nightly, so a model's context window and its price can never drift apart from disagreeing about which upstream snapshot they came from.

Shape

interface ModelCatalogEntry {
  ref: string                  // canonical id: `${provider}/${id}`, e.g. "anthropic/claude-haiku-4-5"
  id: string                   // bare wire id, no provider prefix
  provider: string             // one of the provider registry's ids (see Provider Keys)
  displayName: string
  contextWindow?: number
  maxOutputTokens?: number
  inputCostPer1k?: number
  outputCostPer1k?: number
  supportsFunctionCalling: boolean
  supportsVision: boolean
  judgeCapable: boolean        // see "What makes a model judge-capable" below
  deprecated: boolean          // upstream's deprecation_date has passed
}

Provider coverage

The catalog spans the registry's providers (Anthropic, OpenAI, Gemini, Mistral, OpenRouter, Azure OpenAI, AWS Bedrock, Google Vertex AI, Cohere, Ollama; not yet DeepSeek's own API, whose models appear only under OpenRouter and Ollama) — but "in the catalog" and "routable by Intutic's managed gateway today" are different questions. Only routingLive: true providers (Anthropic, OpenAI, Gemini, Azure OpenAI, AWS Bedrock, Google Vertex AI, Mistral, OpenRouter, DeepSeek as of this writing — see Provider Keys) can be reached through the gateway; Cohere and Ollama are real, browsable catalog entries for providers whose routing is separate, real engineering, not yet built.

Judges are a narrower question still: they run only on self-hosted models. Of the registry, only Ollama is self-hosted (SELF_HOSTED_MODEL_PROVIDER_IDS), and its Ollama Cloud models (ollama/<name>:<size>-cloud) run on ollama.com, so they count as hosted. selfHostedJudgeModelChoices() returns the judge-capable entries left after that filter.

What makes a model judge-capable

judgeCapable is computed once, at generation time, from:

• mode === 'chat' (not an embedding, image, or audio model)
• no audio output
• at least 256 max output tokens
• at least an 8,192-token context window
• not deprecated

This is a floor a judge call can reasonably run under, not a quality ranking — a small, fast, cheap model that clears the floor is judgeCapable, and whether it's a good judge is a separate, workspace-specific decision.

Using it in code

import {
  selfHostedJudgeModelChoices,
  isHostedModelRef,
  findCatalogModel,
  normalizeModelRef,
} from '@intutic/shared-types'

// Judge-capable models a self-hosted judge may use (Ollama, Ollama Cloud excluded)
const choices = selfHostedJudgeModelChoices()

// Refuse a hosted judge model
isHostedModelRef('anthropic/claude-haiku-4-5') // → true
isHostedModelRef('ollama/gpt-oss:120b-cloud')  // → true (Ollama Cloud)
isHostedModelRef('ollama/llama3.1')            // → false
isHostedModelRef('my-org/local-qwen')          // → false (a local alias, not a registry provider)

// Look up a model by its canonical ref or bare id
const entry = findCatalogModel('claude-haiku-4-5')

// Collapse an accidentally-repeated provider prefix
normalizeModelRef('anthropic/anthropic/claude-haiku-4-5')
// → { provider: 'anthropic', id: 'claude-haiku-4-5', ref: 'anthropic/claude-haiku-4-5' }

Custom and BYO model names

A name outside the catalog is legal — an on-prem LiteLLM deployment can serve any model under any alias it chooses. Catalog membership is informational. What is refused is a hosted model: a custom judge reference for which isHostedModelRef() is true.

Related

• Settings & Configuration — Judges
• Provider Keys