Moving reasoning into chat()

Move reasoning, thinking, and thinkingConfig out of modelOptions. Pass reasoning to chat() as a level, or as { level, summary, budgetTokens }.

Upgrade

Delete the reasoning field from modelOptions.
Set reasoning on chat(). Find the old field of your adapter in the table below.
If you set a token budget, move it to budgetTokens. See Budgets.
Run your type checker.

There is no codemod for this change.

// Before
chat({
  adapter: openaiText("gpt-6.1-sol"),
  messages,
  modelOptions: { reasoning: { effort: "high", summary: "auto" } },
});

import { chat } from "@tanstack/ai";
import { openaiText } from "@tanstack/ai-openai";

const messages = [{ role: "user" as const, content: "Hello" }];

// After
chat({
  adapter: openaiText("gpt-6.1-sol"),
  messages,
  threadId: "thread-1",
  reasoning: "high",
});

The thinking text streams by default. To hide it, pass reasoning: { level: "high", summary: false }.

Adapter by adapter

Adapter	Old modelOptions field	Now
openaiText, openaiChatCompletions, azureOpenaiText	reasoning: { effort, summary }	reasoning: "high"
anthropicText	thinking, effort, output_config.effort	reasoning: "high", or { level, budgetTokens } on budget models
geminiText	thinkingConfig	reasoning: "high", or { level, budgetTokens } on Gemini 2.5
geminiTextInteractions	generation_config.thinking_level, thinking_summaries	reasoning: "high"
grokText, openRouterText	reasoning: { effort }	reasoning: "high"
vercelGatewayText, lovableText	reasoning, include_reasoning	reasoning: "high"
bedrockText	reasoning_effort (chat), reasoning (Responses)	reasoning: "high"
groqText, llmGatewayText, cloudflareText	reasoning_effort	reasoning: "high"
byteplusText	thinking, reasoning_effort	reasoning: "high", or "off"
ollamaText	think	reasoning: "high", or "off"
codexText	modelReasoningEffort	reasoning: "high"

"high" is an example. Use any level that your model has. The types show the levels of your model.

Budgets

An Anthropic or Gemini 2.5 budget moves to budgetTokens:

// Before
modelOptions: { thinking: { type: "enabled", budget_tokens: 8000 } }

import { chat } from "@tanstack/ai";
import { anthropicText } from "@tanstack/ai-anthropic";

const messages = [{ role: "user" as const, content: "Hello" }];

// After
chat({
  adapter: anthropicText("claude-haiku-4-5"),
  messages,
  threadId: "thread-1",
  reasoning: { level: "high", budgetTokens: 8000 },
});

The adapter raises max_tokens when it is below the budget.

Azure

azureOpenaiText and createAzureOpenaiText read the levels of the OpenAI model name that you pass, not of your deployment name:

import { chat } from "@tanstack/ai";
import { createAzureOpenaiText } from "@tanstack/ai-openai";

const messages = [{ role: "user" as const, content: "Hello" }];

chat({
  adapter: createAzureOpenaiText("gpt-6.1-sol", "azure-key", {
    resourceName: "my-resource",
    deploymentName: "prod-chat",
  }),
  messages,
  threadId: "thread-1",
  reasoning: "high",
});

If you pass only a deployment name as the model, the adapter has no levels for it, and reasoning is a type error. Pass the OpenAI model name, and set deploymentName.

Codex

modelReasoningEffort stays on the codexText config. It is the default for a call without reasoning.

Removed types

@tanstack/ai-openai: OpenAIReasoningOptions, OpenAIReasoningOptionsWithConcise.
@tanstack/ai-anthropic: AnthropicThinkingOptions, AnthropicAdaptiveThinkingOptions, AnthropicAdaptiveOnlyThinkingOptions, AnthropicAdaptiveOrDisabledThinkingOptions, AnthropicEffortOptions, AnthropicOutputConfigOptions.
@tanstack/ai-gemini: GeminiThinkingOptions.
@tanstack/ai-grok: GrokReasoning, GrokReasoningEffort, GrokBuildProviderOptions.
@tanstack/ai-openrouter: ReasoningOptions. @tanstack/ai-byteplus: BytePlusThinkingOption, BytePlusReasoningEffort.

Done

Your type checker passes, and no modelOptions has a reasoning field. One reasoning option now sets reasoning on every provider. For levels, budgets, and middleware, see Reasoning.