> ## Documentation Index
> Fetch the complete documentation index at: https://vibes-sdk.a7ul.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Thinking / Extended Reasoning

> Enable extended reasoning for Anthropic and Google models by configuring thinking at the model constructor level via provider-specific options.

Extended reasoning lets a model spend additional tokens on internal reasoning before generating a response. This produces more deliberate, multi-step outputs for complex problems - at the cost of higher token usage and latency.

Use extended reasoning for complex multi-step problems, mathematical reasoning, and deep analytical tasks where a more careful thought process improves answer quality.

<Info>
  Vibes does not have a dedicated thinking API. Extended reasoning is configured at the model constructor level using provider-specific options from the Vercel AI SDK. The framework passes these options through automatically.
</Info>

## Anthropic extended thinking

Pass a `thinking` option directly to the `anthropic()` model constructor. Set `budgetTokens` to the maximum number of tokens the model can use for internal reasoning.

```typescript theme={null}
import { anthropic } from "@ai-sdk/anthropic";
import { Agent } from "@vibesjs/sdk";

const model = anthropic("claude-opus-4-5", {
  thinking: { type: "enabled", budgetTokens: 10000 },
});

const agent = new Agent({
  model,
  modelSettings: {
    maxTokens: 16000,  // must exceed thinking.budgetTokens
  },
});

const result = await agent.run(
  "Prove that the square root of 2 is irrational."
);
console.log(result.output);
```

<Warning>
  `maxTokens` in `modelSettings` must be greater than `thinking.budgetTokens`. If `maxTokens` is too low, the API will return an error. For example, if `budgetTokens` is 10000, set `maxTokens` to at least 12000–16000 to leave room for the final response.
</Warning>

## Google extended thinking

Some Gemini models enable extended reasoning by model selection - simply choosing a "thinking" model variant is sufficient. No additional configuration is required.

```typescript theme={null}
import { google } from "@ai-sdk/google";
import { Agent } from "@vibesjs/sdk";

const model = google("gemini-2-5-flash-thinking", {
  // Gemini thinking models enable extended reasoning by model selection.
  // No additional config required for default thinking mode.
});

const agent = new Agent({ model });

const result = await agent.run("Explain the P vs NP problem.");
console.log(result.output);
```

For other Gemini models that support configurable thinking, use the `@ai-sdk/google` provider's constructor options in the same pattern as Anthropic.

## Why is thinking not on AgentOptions?

The framework's `ModelSettings` maps to Vercel AI SDK's top-level `generateText` options - fields like `maxTokens`, `temperature`, and `topP`. These are standard options supported by all providers.

Provider-specific features like `thinking` go through the model constructor's `providerOptions`, which the AI SDK processes before the framework is involved. The Vibes framework does not intercept or re-expose these options - they flow from the model constructor directly to the provider.

This means:

* **Correct:** configure thinking on the model constructor: `anthropic("model", { thinking: ... })`
* **Incorrect:** look for a `thinking` key on `AgentOptions` or `modelSettings` - it does not exist

## API reference

| Configuration             | Location                                                         | Notes                                                         |
| ------------------------- | ---------------------------------------------------------------- | ------------------------------------------------------------- |
| `thinking` option         | Model constructor (e.g. `anthropic("model", { thinking: ... })`) | Provider-specific; not on `AgentOptions`                      |
| `modelSettings.maxTokens` | `AgentOptions.modelSettings` or run-time `modelSettings`         | Must exceed `budgetTokens` when thinking is enabled           |
| `thinking.type`           | `"enabled"`                                                      | Required field for Anthropic thinking                         |
| `thinking.budgetTokens`   | `number`                                                         | Max tokens for internal reasoning; `maxTokens` must be higher |

***

<CardGroup cols={2}>
  <Card title="Models" icon="brain" href="/concepts/models">
    Configure models, providers, and model settings
  </Card>

  <Card title="Anthropic AI SDK Thinking Docs" icon="arrow-up-right-from-square" href="https://sdk.vercel.ai/providers/ai-sdk-providers/anthropic#thinking">
    Provider-level thinking configuration reference
  </Card>
</CardGroup>