Parameters and fallback

The SDK checks sampling and token options for recognized OpenAI model families, including GPT-6. With reasoning enabled, an unsupported temperature is omitted with a warning. Reasoning is not silently disabled to accept a temperature value.

import { generateText } from "@yourgpt/llm-sdk";
import { openai } from "@yourgpt/llm-sdk/openai";

const result = await generateText({
  model: openai("gpt-6-luna", {
    onWarning: (warning) => console.warn(warning),
  }),
  prompt: "Reply with exactly OK.",
  temperature: 0, // Omitted when the model's default reasoning disallows it.
});

If the model supports non-reasoning mode and your application needs sampling control, explicitly set reasoningEffort: "none" on the generation request. The SDK then retains temperature. Unsupported reasoning settings fail before the provider call. Runtime requests can override the setting through config.reasoningEffort.

Endpoint selection

On OpenAI's endpoint, recognized GPT-5.4+ and GPT-6 reasoning requests with tools use Responses automatically. You can also select Responses explicitly. Custom OpenAI-compatible endpoints keep Chat Completions routing unless you opt in. parameterPolicy: "passthrough" disables normalization for an endpoint with different rules; it can cause provider errors if used with unsupported settings.

Recover from a specific parameter error

Compatibility errors stop by default. fallbackErrors lets your application switch models for a specific error without retrying the same invalid request:

import { createFallbackChain } from "@yourgpt/llm-sdk/fallback";
import { createOpenAI } from "@yourgpt/llm-sdk/openai";

const provider = createOpenAI();
const adapter = createFallbackChain({
  models: [
    provider.languageModel("gpt-6-luna"),
    provider.languageModel("gpt-4o"),
  ],
  retries: 2,
  fallbackErrors(error) {
    const e = error as { status?: number; code?: string; param?: string };
    return (
      e?.status === 400 &&
      ["unsupported_parameter", "unsupported_value"].includes(e.code ?? "") &&
      e.param === "temperature"
    );
  },
});

Pass the returned adapter to createRuntime({ adapter }). Transient failures retain their configured retries. Do not match every HTTP 400 or mark compatibility errors retryable. Cancellation never triggers a fallback, and a stream cannot switch models after it has emitted content or tool events.

Guardrail decisions

TRAINING_CONFLICT belongs to your application's policy. An API failure is not a successful classification and must not silently become a respond decision. Handle an exhausted fallback chain through your application's error or review path. Native Decisions is a separate API; it does not automatically replace existing guardrails or retry failed decisions.

The Feature Lab tests both normal parameter filtering and a deliberately invalid request with one explicit fallback. See the versions and compatibility for release availability.