ByteDance Doubao

Direct access versus this gateway

Ark's direct data-plane base is https://ark.cn-beijing.volces.com/api/v3 and uses a project API key. Its control plane uses ark.cn-beijing.volcengineapi.com and is not the inference address. Gateway calls use this site's base URL and Bearer token, and this page currently declares only /v1/chat/completions. Ark recommends evaluating Responses for new direct integrations, but that does not expose the same gateway route.

Model ID, endpoint ID, and gateway model name

A console display name is for reading. Preset model calls normally use a model ID, such as doubao-seed-2-1-pro-260628 in the official quickstart; a custom inference deployment uses an endpoint ID. Gateway clients use a name enabled in the gateway model list because administrators may map it to either official identifier. Do not copy these three values interchangeably.

Authentication, projects, and protocol boundaries

An Ark API key belongs to one project and can be restricted by model, custom endpoint, and source IP. Direct OpenAI-compatible extensions often go in extra_body, while capabilities such as embeddings may require a native protocol. Media generation has separate task workflows and cannot be called with Chat merely because it shares the Doubao brand.

Streaming, tools, and errors

Official Chat and Responses use different SSE event shapes; a Responses stream cannot be parsed as Chat delta. Web search, knowledge bases, MCP, and function tools are model- and protocol-specific, and applications must validate tool arguments and side effects. A stream may carry an error after HTTP 200, so keep reading error.type and the request ID.

Cost safety and production checklist

Coding Plan uses a separate base URL, and the pay-as-you-go address can create out-of-plan charges. Set output caps, concurrency, timeouts, bounded retries, and account alerts, and verify tool, cache, and media pricing. Before launch, check region, project, key scope, model or endpoint identifier, protocol support, moderation, and current pricing.

Online, Fast, batch, and billing

Model variants and service tiers are separate choices: concurrent online calls are not batch inference, and Turbo is not Fast. The gateway's billing adaptation can represent scenario prices, but configuring them does not add batch submission endpoints. Verify routes and channel capabilities independently. Audio-cache overlap needs deduplication, and storage needs measured duration. The Doubao billing guide covers the pricing form, length bands, and reconciliation examples.