Skip to main content
These documentation defaults were checked against provider catalogs on October 4, 2026. Agentium’s release version and a provider’s model version are independent.

​
Text models used in examples

These selections follow the providers’ published model catalogs: OpenAI, Anthropic, and Google. Check account access, region, and pricing before deploying. The environment names above are conventions used by the example code, not automatic Agentium configuration.

​
Keep the endpoint and parameters compatible

  • OpenAI: v4 routes GPT-6 function-tool calls through Responses. It drops unsupported sampling fields for reasoning models. Use a supported effort; Astra and Sol do not offer none, while Luna does. A custom OpenAI-compatible gateway must implement the endpoint the adapter selects.
  • Anthropic: v4 uses adaptive thinking for Claude 5 when reasoning is enabled. Use effort for these examples; an older budgetTokens recipe does not set a fixed thinking budget on the adaptive path.
  • Gemini: v4 maps Gemini 3 reasoning effort to thinking level and removes sampling fields for Gemini 3.8. A Gemini 2.5 token-budget example is not the same configuration.
  • Provider continuation: retain opaque provider history for a tool loop. Switching provider/model may require a new session; do not reuse signed reasoning or tool continuation from a different provider.

​
Specialized adapters have their own versions

Speech models must match the adapter’s endpoint and accepted model families. In particular, ElevenLabs stream-input synthesis uses its supported v2 families; replacing that ID with a newer dialogue model would break the example. Embedding dimensions, rerank models, image models, and device models also have independent contracts. Azure deployment names, Bedrock inference profiles, Vertex regional availability, and third-party gateway IDs remain host configuration. Do not replace them mechanically with a direct API model ID. Historical migration examples and pricing fixtures retain their historical names where they explain old behavior.

​
Update cost accounting too

The v4 built-in cost table predates the current GPT-6 examples. Supply matching CostTracker.pricing entries when changing models; the cost recipes include dated text-rate overrides. Token usage can be correct while an outdated or missing price makes the dollar estimate wrong.

​
What has been checked

Complete examples are type-checked against Agentium v4 declarations. Harness recipes run with deterministic providers to verify tool loops, context, session ownership, events, and cleanup. These checks do not certify live provider behavior or your account’s access. Run a small application evaluation before changing the model in a deployed service.