Models and Registry

Models are defined in src/facade/ml/models.generated.ts and loaded into an in-memory registry in src/facade/ml/models.ts.

Model Structure

Each Model<TApi> (from src/facade/ml/types.ts) includes:

• id and name
• api and provider
• baseUrl
• reasoning (boolean support flag)
• input types — an array of "text" and/or "image"
• cost per million tokens: input, output, cacheRead, cacheWrite
• contextWindow and maxTokens
• optional headers
• optional compat (only for openai-completions / openai-responses APIs)

Registry Functions

• getModel(provider, modelId) returns a single model, throwing if unknown.
• tryGetModel(provider, modelId) returns the model or undefined.
• getModels(provider) returns all models for a provider.
• getProviders() returns the catalog's provider names.
• findModels(filters?) filters by provider, api, reasoning, supportsImageInput, and nameIncludes.
• registerCustomModel(model) adds one model; loadCustomModels(models) adds many.

These delegate to the shared modelRegistry singleton (an instance of the exported ModelRegistry class). You can also construct your own ModelRegistry with a custom seed.

import { findModels } from "indusagi/ai";

const reasoningModels = findModels({ reasoning: true, supportsImageInput: true });

Cost Calculation

• estimateCost(model, { inputTokens, outputTokens, cacheReadTokens?, cacheWriteTokens? }) returns a { input, output, cacheRead, cacheWrite, total } breakdown.
• calculateCost(model, usage) mutates and returns usage.cost, computing each component from the per-million model.cost rates.

Cache reads and cache writes are priced separately from input and output tokens.

Reasoning Levels

• supportsXhigh(model) returns true only for gpt-5.1-codex-max, gpt-5.2, and gpt-5.2-codex.
• Providers that do not support "xhigh" clamp it to "high" (see clampReasoning in Streaming Model).

Comparing Models

modelsAreEqual(a, b) returns true when both models share the same id and provider (and both are non-null).

Compatibility Overrides

Model.compat overrides endpoint behavior for OpenAI-compatible providers. It is typed as OpenAICompletionsCompat for the openai-completions API and OpenAIResponsesCompat for openai-responses. See AI Providers for the full field list.