​

Compatibility and limits

Use this page when you need a quick answer about whether a specific OpenAI or Anthropic feature works on Radium.

​

Supported and unsupported endpoints

​

Parameter limits

​

Tool calling

Tool calling works on hal-1.0, clarke-1.0, and tycho-1.0 through both the OpenAI and Anthropic endpoints. The schema format, tool_choice behavior, and parallel tool execution match the OpenAI specification. There is no extra latency for tool-enabled requests.

​

Structured outputs

JSON mode (response_format: {type: "json_object"}) and JSON Schema (response_format: {type: "json_schema", ...}) are supported on hal-1.0 and clarke-1.0. Tycho 1.0 supports JSON mode. Structured outputs follow the same rules as OpenAI: the model may refuse if the prompt cannot satisfy the schema, and a refusal field may be present.

​

Prompt caching

The cache_control parameter is supported on all models. Cache hits are billed at the cached input rate. There is no additional latency for writing or reading from cache.

​

Vision limits

Vision input is supported on hal-1.0 and clarke-1.0. Images can be sent as base64-encoded data URLs or public HTTP(S) URLs. See the Multimodal images example for code samples. Tycho 1.0 does not support vision input.

​

Rate limits

Rate limits are enforced at the organization level and shared across all API keys. See Rate limits for full details on error codes and backoff strategies.

​

SDK compatibility

Any SDK or framework that can target the OpenAI Chat Completions API or the Anthropic Messages API works with Radium after changing the base URL and API key. Official SDKs, Claude Code, LangChain, LangGraph, Vercel AI SDK, and others have been verified.