Skip to main content
Different providers express reasoning in different ways (token budgets, effort levels, or model-specific flags). Hebo Gateway standardizes this into a single reasoning configuration so you can keep one client and switch models safely.

​
Reasoning Effort

You can pass a normalized reasoningEffort parameter with a consistent shape across providers. It supports the following values:
ValueMeaning
noneDisable reasoning.
minimalLowest reasoning effort.
lowLow reasoning effort.
mediumDefault reasoning effort.
highHigh reasoning effort.
xhighMaximum reasoning effort.
Hebo automatically maps this setting to the appropriate provider-specific controls (effort tiers, token budgets, or internal flags), so your application code remains unchanged when switching models.

​
Basic Example

Vercel AI SDK
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import { generateText } from "ai";

const hebo = createOpenAICompatible({
  name: "hebo",
  apiKey: process.env.HEBO_API_KEY,
  baseURL: "https://gateway.hebo.ai/v1",
});

const result = await generateText({
  model: hebo("openai/gpt-oss-20b"),
  prompt: "Why did the monkey bring a ladder to the bar?",
  providerOptions: {
    hebo: {
      reasoningEffort: "medium",
    },
  },
});

console.log("Reasoning: ", result.reasoningText);
console.log("Text: ", result.text);
Some models don’t support returning reasoning content, in which case the result.reasoning parameter will be empty. The model still used reasoniong internally.

​
Chain-Of-Thoughts

If you use streaming, you can see the thought process of the model in realtime.
Vercel AI SDK
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import { streamText } from "ai";

const hebo = createOpenAICompatible({
  name: "hebo",
  apiKey: process.env.HEBO_API_KEY,
  baseURL: "https://gateway.hebo.ai/v1",
});

const result = streamText({
  model: hebo("openai/gpt-oss-20b"),
  prompt: "Tell me a very long story about monkeys",
  providerOptions: {
    hebo: {
      reasoningEffort: "medium",
    },
  },
});

for await (const part of result.fullStream) {
  if (part.type === "reasoning-delta") {
    process.stdout.write("<REASONING STEP>\n");
    process.stdout.write(part.text);
    process.stdout.write("</REASONING STEP>\n\n");
  }

  if (part.type === "text-delta") {
    process.stdout.write(part.text);
  }
}
Not all models provide the detailed chain-of-thoughts output, summ will only provide a reasoning summary at the end.

​
Follow-up Calls

Follow-up calls should reuse the previous response messages, so the next request preserves the model’s context, including any tool calls and internal thought process. Some providers also include reasoning metadata that must be passed along to keep model behavior consistent across a conversation. Hebo Gateway handles this automatically, as long as you include the full returned messages object in your next request. 
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import { generateText } from "ai";

const hebo = createOpenAICompatible({
  name: "hebo",
  apiKey: process.env.HEBO_API_KEY,
  baseURL: "https://gateway.hebo.ai/v1",
});

const messages = [
  { role: "user", content: "Why did the monkey bring a ladder to the bar?" } as const,
];

const first = await generateText({
  model: hebo("openai/gpt-oss-20b"),
  messages,
  providerOptions: {
    hebo: {
      reasoningEffort: "medium",
    },
  },
});

console.log(first.text + "\n");

messages.push(...first.response.messages);
messages.push({
  role: "user",
  content: "Why did the monkey bring a ladder to the bar?",
} as const);

const second = await generateText({
  model: hebo("openai/gpt-oss-20b"),
  messages,
  providerOptions: {
    hebo: {
      reasoningEffort: "medium",
    },
  },
});

console.log(second.text);
Internally, this automatically preserves reasoning details and thought signatures across follow-up calls.

​
Advanced Parameters

For more configurability, the normalized reasoning object supports the following parameters: 
"reasoning": {
  "enabled": true,      // turn reasoning on/off
  "effort": "medium",   // none | minimal | low | medium | high | xhigh
  "maxTokens": 2048,   // max tokens for reasoning (advanced)
  "exclude": false      // exclude reasoning from response payload
}
Keep in mind, when you tweak the individual parameters that some models might not support your specific settings.