Interface: IModelCaps

Defined in: internal/model-registry/types.ts:217

Capabilities and cost of a model. Checked before building a request.

Properties

contextTokens

contextTokens: number

Defined in: internal/model-registry/types.ts:240

Context window in tokens.

---

costCacheReadPerMTokUsd?

optional costCacheReadPerMTokUsd?: number

Defined in: internal/model-registry/types.ts:256

USD per 1M input tokens read from the provider's prompt cache.

Prices the cachedInputTokens of a call.

Example

0.1 - claude-haiku-4-5 (0.1 × its 1.00 input price)

Default

Anthropic: 0.1 × costInPerMTokUsd; any other provider: costInPerMTokUsd (the rate differs per OpenAI model: set it)

---

costCacheWritePerMTokUsd?

optional costCacheWritePerMTokUsd?: number

Defined in: internal/model-registry/types.ts:266

USD per 1M input tokens written to the provider's prompt cache.

Prices the cacheWriteInputTokens of a call.

Example

1.25 - claude-haiku-4-5, 5-minute cache (1.25 × its 1.00 input price)

Default

Anthropic: 1.25 × costInPerMTokUsd; any other provider: costInPerMTokUsd (OpenAI bills no write)

---

costInPerMTokUsd

costInPerMTokUsd: number

Defined in: internal/model-registry/types.ts:244

USD per 1M input tokens.

---

costOutPerMTokUsd

costOutPerMTokUsd: number

Defined in: internal/model-registry/types.ts:246

USD per 1M output tokens.

---

embeddings

embeddings: boolean

Defined in: internal/model-registry/types.ts:238

True only for embedding models.

---

images

images: boolean

Defined in: internal/model-registry/types.ts:221

Accepts image parts.

---

maxOutputTokens

maxOutputTokens: number

Defined in: internal/model-registry/types.ts:242

Max output tokens per call.

---

pdf

pdf: boolean

Defined in: internal/model-registry/types.ts:223

Accepts PDF parts.

---

reasoning

reasoning: boolean

Defined in: internal/model-registry/types.ts:227

Exposes reasoning control.

---

structuredOutput

structuredOutput: boolean

Defined in: internal/model-registry/types.ts:225

Supports schema-constrained JSON output.

---

temperature

temperature: boolean

Defined in: internal/model-registry/types.ts:236

Accepts a sampling temperature.

Decides whether a call may carry temperature; the gateway refuses one on a model without this cap (400) instead of letting the provider drop it with a warning.

False for the OpenAI reasoning models (o-series, GPT-5 family): they reject the parameter. True for the Anthropic models.

---

tools

tools: boolean

Defined in: internal/model-registry/types.ts:219

Can call tools.