aidy-models

aidy-models is a model registry for LLM providers and models.

It provides a models.json file and keeps the data shaped for runtime consumers:

• provider defaults
• model-level overrides
• capability metadata
• pricing metadata
• runtime compatibility metadata

Agent Skill

Install the Model Pricing skill to help agents answer pricing questions from the latest registry data:

npx skills add ImSingee/aidy-models

Data Sources

The current generator merges:

• models.dev
• lobehub

Generate

bun run generate

This rewrites models.json.

Output Shape

models.json has the following top-level structure:

{
  "_meta": {},
  "providers": {},
  "models": {}
}

_meta

Generation metadata:

• generatedAt: ISO timestamp of the last generation
• sources.modelsDev: provider and model counts from models.dev
• sources.lobehub: provider and model counts from Lobehub, plus commitHash
• merged: counts after merge and normalization
• overridesApplied: number of local overrides applied from src/generate/overrides.ts

providers

providers is a map keyed by canonical provider id.

Each provider entry contains default runtime metadata for that provider.

models

models is a map keyed by canonical provider id.

Each value is an array of models for that provider.

Model entries may override provider defaults for cases where a single provider serves multiple protocols or base URLs.

Merge Semantics

Consumers should treat provider-level fields as defaults and model-level fields as overrides.

Recommended precedence:

1. Start from provider
2. Apply model
3. For nested compat, merge provider.compat first and then model.compat

This matters for fields such as:

• api
• baseUrl
• headers
• compat

Example:

• github-copilot has provider-level static headers
• some github-copilot models override api from openai-completions to openai-responses or anthropic-messages
• opencode provider defaults to openai-completions, while some models override to openai-responses, anthropic-messages, or google-generative-ai

Provider Fields

Provider describes provider defaults.

Core runtime fields

• id: canonical provider id
• name: display name
• api: default API protocol for the provider
• baseUrl: default base URL for the provider
• headers: static default headers to send for this provider
• compat: provider-level runtime compatibility defaults

Metadata fields

• official: whether the provider is treated as an official first-party source
• featured: whether the provider should be highlighted in consumer UIs
• description: provider description
• url: provider homepage
• doc: docs URL
• enabled: upstream enablement flag if available
• checkModel: model id suitable for health checks or API verification
• apiKeyUrl: API key management page

Model Fields

Model describes a concrete model entry.

Identity and runtime fields

• id: model id used by the provider API
• name: display name
• api: optional model-level API override
• baseUrl: optional model-level base URL override
• headers: optional model-level static headers
• compat: optional model-level compatibility overrides
• _: optional unstable passthrough metadata

Descriptive fields

• description: model description
• type: upstream type such as chat
• family: model family
• releasedAt: release date string
• knowledge: knowledge cutoff string if available
• openWeights: whether the model has open weights
• deprecated: whether the model is deprecated upstream

Capability fields

• abilities.toolCall: supports tool / function calling
• abilities.reasoning: supports reasoning mode
• abilities.vision: accepts image input
• abilities.structuredOutput: supports structured output
• abilities.search: supports native search
• abilities.imageOutput: supports image generation
• abilities.video: supports video input or output, depending on source
• abilities.attachment: supports file or attachment input
• abilities.temperature: supports configurable temperature
• abilities.interleaved: supports interleaved thinking / tool behavior

Token and modality fields

• contextWindow: max input context
• maxOutput: max output tokens
• modalities.input: accepted input modalities
• modalities.output: supported output modalities

_ metadata

_ is reserved for unstable local metadata that is useful to runtime consumers but not normalized as first-class schema yet.

OpenAI reasoning models may include:

• _.reasoningEffort.enum: supported reasoning effort values
• _.reasoningEffort.default: default reasoning effort value

reasoningEffort constrains effort on reasoning for reasoning models. Currently supported values are none, minimal, low, medium, high, and xhigh.

Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response.

Pricing fields

• pricing.currency: pricing currency, such as USD or CNY
• pricing.unit: billing unit used by this pricing definition
• pricing.basePricing: base rates by pricing target
• pricing.adjustments: conditional pricing adjustments

Pricing Schema

pricing contains:

• unit: billing denominator such as millionTokens or image
• basePricing: default rates keyed by pricing target such as textInput
• adjustments: conditional changes applied on top of basePricing

The current schema assumes a single unit per model pricing definition.

Each adjustment contains:

• mode: multiplier or absolute
• when: condition map
• unless: optional exception condition map, or a list of exception condition maps
• values: target -> factor or final rate

Condition values may be:

• string, number, or boolean for a matching condition
• [number, number | "infinity"] for ranges such as textTotalInput: [0.2, "infinity"]

Adjustment application order:

1. Start from basePricing
2. Iterate over adjustments in array order
3. For each adjustment, first match when
4. Skip the adjustment if any unless clause also matches
5. For each matching target:
   • if mode is absolute, replace the current rate
   • if mode is multiplier, multiply the current rate

If unless is a single object, all keys in that object must match to suppress the adjustment. If unless is an array, matching any one entry suppresses the adjustment.

This means adjustment order is significant:

• absolute followed by multiplier: both apply, because the later multiplier multiplies the overridden rate
• multiplier followed by absolute: only the later absolute result remains, because it replaces the already-multiplied rate

Pricing target enum

key	Meaning
textInput	prompt or other input text billing
textOutput	generated text billing
textInput_cacheRead	prompt cache read / cache hit billing
textInput_cacheWrite	prompt cache write billing
audioInput	audio input billing
audioOutput	audio output billing
audioInput_cacheRead	cached audio input read billing
imageInput	image input billing
imageInput_cacheRead	cached image input read billing
imageOutput	image output billing
imageGeneration	image generation billing
videoGeneration	video generation billing

unit enum

unit	Meaning
millionTokens	price per 1,000,000 tokens
millionCharacters	price per 1,000,000 characters
image	price per image
megapixel	price per megapixel
second	price per second

Use the billing denominator that matches the provider's published pricing. For example:

• text and most chat models use millionTokens
• some TTS models bill textInput by millionCharacters
• image generation may bill by image or megapixel
• audio or video duration-based pricing may bill by second

when key enum

The when object uses condition keys to describe when an adjustment applies.

key	Value type	Meaning
cacheTtl	string	prompt cache TTL such as 5m, 1h, or 24h
fastMode	boolean	whether provider fast mode is enabled
serviceTier	"flex" | "priority"	OpenAI service tier selector
textTotalInput	[number, number | "infinity"]	total input-token bucket, including textInput + textInput_cacheRead + textInput_cacheWrite, using pricing.unit as the denominator
textOutput	[number, number | "infinity"]	output-token bucket, using pricing.unit as the denominator
quality	string	image quality variant such as standard or hd
size	string	image size such as 1024x1024
generateAudio	boolean	whether audio generation is enabled
thinkingMode	boolean	whether a model is in thinking / reasoning mode

Other condition keys may appear if upstream pricing introduces more dimensions. When that happens, the value type still follows PricingConditionValue.

serviceTier is currently intended for OpenAI-style responses runtimes. The known values are:

• flex: lower-cost background or latency-tolerant service tier
• priority: premium higher-priority service tier

values key enum

The keys in basePricing and adjustments.values use the same enum:

key	Meaning
textInput	prompt or other input text billing
textOutput	generated text billing
textInput_cacheRead	prompt cache read / cache hit billing
textInput_cacheWrite	prompt cache write billing
audioInput	audio input billing
audioOutput	audio output billing
audioInput_cacheRead	cached audio input read billing
imageInput	image input billing
imageInput_cacheRead	cached image input read billing
imageOutput	image output billing
imageGeneration	image generation billing
videoGeneration	video generation billing

values means different things depending on mode:

• if mode is multiplier, values[target] is a factor such as 1.5 or 2
• if mode is absolute, values[target] is the final rate in pricing.unit

Example: multiplier

{
  "currency": "USD",
  "unit": "millionTokens",
  "basePricing": {
    "textInput": 3,
    "textOutput": 15,
    "textInput_cacheRead": 0.3,
    "textInput_cacheWrite": 3.75
  },
  "adjustments": [
    {
      "mode": "multiplier",
      "when": {
        "textTotalInput": [0.272, "infinity"]
      },
      "values": {
        "textInput": 2,
        "textOutput": 1.5,
        "textInput_cacheRead": 2
      }
    },
    {
      "mode": "multiplier",
      "when": {
        "serviceTier": "priority"
      },
      "values": {
        "textInput": 2,
        "textOutput": 2,
        "textInput_cacheRead": 2
      }
    },
    {
      "mode": "multiplier",
      "when": {
        "cacheTtl": "1h"
      },
      "values": {
        "textInput_cacheWrite": 1.6
      }
    }
  ]
}

Example: absolute

{
  "currency": "USD",
  "unit": "image",
  "basePricing": {
    "imageGeneration": 0.04
  },
  "adjustments": [
    {
      "mode": "absolute",
      "when": {
        "quality": "hd",
        "size": "1024x1024"
      },
      "values": {
        "imageGeneration": 0.08
      }
    }
  ]
}

Compat Schema

compat stores runtime behavior that is not captured by basic fields like api, baseUrl, or abilities.

The goal is to move provider- and model-specific runtime quirks out of client code and into registry data.

compat is split by runtime family:

• openaiCompletions
• openaiResponses
• anthropic
• bedrock
• google
• googleGeminiCli

openaiCompletions

For OpenAI-compatible chat-completions style APIs.

• supportsStore
• supportsDeveloperRole
• supportsReasoningEffort
• supportsUsageInStreaming
• maxTokensField: max_completion_tokens or max_tokens
• requiresToolResultName
• requiresAssistantAfterToolResult
• requiresThinkingAsText
• requiresMistralToolIds
• thinkingFormat: openai, zai, or qwen
• openRouterRouting
• vercelGatewayRouting
• supportsStrictMode
• assistantContentFormat: string or parts
• toolCallIdStrategy: preserve, openai-40, pipe-call-40, or mistral-9

openaiResponses

For OpenAI Responses-style APIs.

• toolCallIdStrategy: preserve or responses-fc64
• longPromptCacheTtl: currently 24h
• supportsServiceTier
• supportsAdditionalServiceTiers: array of flex or priority

anthropic

• longPromptCacheTtl: currently 1h
• supportsFastMode
• supportsAdaptiveThinking
• xHighReasoningEffort: high or max

bedrock

• supportsAdaptiveThinking
• supportsPromptCaching
• supportsThinkingSignature
• xHighReasoningEffort

google

For Google Gemini / Vertex style reasoning and tool behavior.

• requiresToolCallId
• supportsMultimodalFunctionResponse
• reasoningMode: level or budget
• reasoningLevelMap
• defaultThinkingBudgets

googleGeminiCli

For Cloud Code Assist / Gemini CLI style APIs.

• toolSchemaFormat: parameters or input_schema
• reasoningMode
• reasoningLevelMap
• defaultThinkingBudgets

Canonical API Values

Current api values used in this registry include:

• openai-completions
• openai-responses
• azure-openai-responses
• anthropic-messages
• bedrock-converse-stream
• google-generative-ai
• google-gemini-cli
• google-vertex

Notes

• Not every field is present on every provider or model.
• Some upstream fields and local override metadata that are useful but not yet normalized may live under _. These fields are unstable and may change at any time.