kinds | What kind of model it is: chat, image, video, speech, transcription, embeddings or moderation. A chat model that can also answer with pictures is chat. Filter on this field. |
primary_operation | The operation the model is chiefly for, as the catalog states it, or null when no provider serves that operation right now. |
operations | What the model does, one entry per job that the catalog states and a provider serves right now. operation is one of chat, image_generation, image_edit, video_generation, speech, transcription, translation, embeddings and moderation; translation is transcription into English. kind is the job’s kind. |
operations[].endpoints | Every endpoint name that serves the job. chat lists all three chat endpoints: chat.completions, messages and responses. Other values are embeddings, moderations, images.generations, images.edits, videos.generations, audio.speech, audio.transcriptions and audio.translations. |
operations[].limits | How large one request may be. context_window on chat is the token budget for prompt and answer together; max_images on image_generation and image_edit is the largest n. max_wait_seconds on video_generation is how long a job may run before it fails with timeout. max_output_megapixels on video_generation is the largest frame a video model delivers, in megapixels of 1,048,576 pixels. context_window on embeddings and moderation is the token budget of each input, and max_inputs is how many inputs one request may carry. max_input_characters on speech caps input. max_file_bytes on transcription and translation caps the audio file. max_documents on rerank caps documents. A request above max_inputs, max_input_characters, max_file_bytes or max_documents gets 400 invalid_request_error that names the limit. A limit the catalog does not record is omitted. |
operations[].parameters | Optional request fields the job accepts, keyed by name. Each value describes the field in a subset of JSON Schema (type, enum, default, minimum, maximum, maxItems, maxLength, items, properties) plus x-support. x-support is native when NagaAI passes the field as sent. It is limited when the field takes only the values the other keys allow, such as "maxItems": 1 on stop. The values include those NagaAI converts for the provider. A reasoning_effort level the provider lacks goes as its nearest level. A JSON schema in response_format goes into the prompt. Names come from a fixed list per job: the fields NagaAI checks when it picks a provider. A name appears when at least one provider of the model accepts it, and the description accepts every value at least one of them accepts. NagaAI sends a value only to a provider that accepts it, see Differences from OpenAI and Anthropic. |
pricing | One row per meter. price is credits per unit as a decimal string, where one credit is one US dollar: "0.000001" per input_token is $1 per million tokens. meter names what NagaAI counts, such as input_text_token, output_audio_hour, output_video_second or output_video_megapixel_second; input_token, output_token and token price whatever has no row of its own. meter request with unit request is a fee one request pays once, on top of its other rows. conditions lists the request values a row applies to, such as resolution on a video row, and is empty when the row always applies. A row with operation in conditions applies to that job of the model only, such as image_generation, and there it replaces the row of the same meter without it. minimum, when present, is the least one request costs under the row, however few units it uses. max_price, when present, is the most one unit costs. The rate then depends on the provider that serves the request, see Limit the price. |
architecture.input_modalities | What the model accepts: text, image, audio, file. A request with another input type fails with 400 before it reaches a provider. |
aliases | Other names the model answers to. The first is the bare model name. |
created, description, owned_by | When the model joined the catalog, a short description, and the publisher’s slug, the part of id before the slash (or null). |
deprecates_at | Unix time after which the model returns 410 model_deprecated. |
precision | The weight precisions the model’s providers publish they serve it in, such as ["bf16", "fp8"], each once and in alphabetical order. Absent when no provider publishes one. Not every provider publishes it, so a request without routing.precision may be served in a precision not listed here. |