From Anthropic
Radium serves the Anthropic Messages endpoint atPOST /v1/messages. The Anthropic SDK, Claude Code, PydanticAI, and any framework that targets Anthropic can point at Radium by changing the base URL.
Before and after
Python
TypeScript
Model mapping
Claude Code setup
Set four environment variables before runningclaude:
ANTHROPIC_BASE_URL points Claude Code at Radium instead of Anthropic. ANTHROPIC_AUTH_TOKEN is your Radium API key. ANTHROPIC_MODEL is the default model for long reasoning tasks. ANTHROPIC_SMALL_FAST_MODEL is used for summaries, background tasks, and quick completions.
Supported endpoints
Reading the response
When reasoning is enabled, the response contains athinking content block before the text block. Read the last content block, not the first one. This is the same pattern the Anthropic SDK uses.
Prompt caching
Thecache_control parameter is supported on all models. Cache hits are billed at the cached input rate: 10% of the input price for Hal 1.0 and Clarke 1.0, 20% for Tycho 1.0.
Known differences
- Model strings are different. Using
claude-*returns a model-not-found error. - Radium does not publish the names of underlying open-weight models.
- Rate limits and pricing differ. See Rate limits and Pricing.