Protocol and SDK support

Check each model's protocols, capabilities, streaming support, and rates in the model catalog, then call GET /v1/models with your credential to confirm accessible models. The catalogs serve different purposes: GET /api/v1/public/models supplies public capability details; /v1/models returns OpenAI-compatible object and data, with id, object, created, and owned_by on model entries. It does not return protocols. Never infer protocol support from a model name.

Use an AICredits API key or a short-lived Embedded token for the data API. Keys issued by other model services cannot authenticate to the platform. The official SDKs use Bearer; compatible SDKs can carry the same platform credential as shown below.

protocols value	Public operation	Compatible SDK / authentication
openai_chat_completions	POST /v1/chat/completions	OpenAI SDK; Bearer
openai_responses	POST /v1/responses	OpenAI SDK; Bearer
openai_embeddings	POST /v1/embeddings	OpenAI SDK; Bearer
openai_images	POST /v1/images/generations	OpenAI SDK; Bearer
anthropic_messages	POST /v1/messages	Anthropic SDK; x-api-key or Bearer
gemini	POST /v1beta/models/{model}:generateContent	Gemini SDK; x-goog-api-key or Bearer
videos	POST /v1/videos/generations	HTTP / official gateway SDK; Bearer

Gemini ?key= authentication is accepted only under /v1beta; prefer headers to keep credentials out of URL logs. SDK badges indicate protocol compatibility, not endorsement or identical parameters, tool behavior, or outputs across models.

Protocol compatibility across models

• OpenAI and Claude: alongside their familiar request formats, the platform provides Messages compatibility for OpenAI models and Chat Completions / Responses compatibility for Claude. Use only protocols declared in the public catalog.
• Gemini: native generation, streaming, token counting, and model lookup, plus Messages / Chat compatibility paths. See Gemini integration.
• Grok: Chat, Responses, and Messages compatibility, plus image and video generation for configured models.
• DeepSeek, Kimi and MiniMax: Chat / Messages / Responses compatibility paths are implemented; choose published models from the catalog.
• Zhipu GLM: Chat / Messages paths are implemented. Responses is not implemented.

The Chinese model guide covers discovery, reasoning options, multi-turn tools and streaming for all four families. SDK listPublicModels returns public protocols and rates; intersect exact IDs with authenticated listModels before choosing a model.

Implemented interfaces still require current plan entitlements, enabled models, and applicable rates. Request-time authorization is authoritative; a catalog result is not an availability guarantee.

Token counting and model lookup

Capability	Public operation	Usage
Messages token counting	POST /v1/messages/count_tokens	Supports enabled OpenAI, Grok, and Claude paths; Grok uses a local estimate, which cannot replace final billing details
Gemini model catalog	GET /v1beta/models, GET /v1beta/models/{model}	Returns currently entitled and callable Gemini models
Gemini token counting	POST /v1beta/models/{model}:countTokens	Uses the native Gemini request shape; permissions and model configuration still apply

Token counting does not run generation. Platform records determine actual usage and cost after generation. See the quickstart for examples.

SDK coverage and limits

Official SDK 0.3.0 provides TypeScript, Swift, and Kotlin gateway wrappers, plus Embedded balance, transaction, purchase, and support APIs. TypeScript supports Chat, Responses, Messages, and Gemini SSE; Swift/Kotlin currently provide non-streaming JSON calls.

• Responses HTTP requests do not support previous_response_id; pass conversation context explicitly. WebSocket / Realtime are outside the current SDK scope.
• Image generation accepts JSON; multipart edits and streaming image events are not supported.
• Video edits, status lookup, content download, asynchronous polling, and delayed settlement are not available. A pending response is not a completed asset or a final charge.
• Ollama Cloud is still in development. Subscription OAuth is not publicly available.

Stream endings differ by protocol; see streaming and cancellation.