Migrate AI SDK 5.x to 6.0 Beta
AI SDK 6 is currently in beta and introduces new capabilities like agents and tool approval. This guide will help you migrate from AI SDK 5.0 to 6.0 Beta. Note that you may want to wait until the stable release for production projects. See the AI SDK 6 Beta announcement for more details on what's new.
Recommended Migration Process
- Backup your project. If you use a versioning control system, make sure all previous versions are committed.
- Upgrade to AI SDK 6.0 Beta.
- Follow the breaking changes guide below.
- Verify your project is working as expected.
- Commit your changes.
AI SDK 6.0 Beta Package Versions
You need to update the following packages to the beta versions in your package.json file(s):
aipackage:6.0.0-beta(or use the@betadist-tag)@ai-sdk/providerpackage:3.0.0-beta(or use the@betadist-tag)@ai-sdk/provider-utilspackage:4.0.0-beta(or use the@betadist-tag)@ai-sdk/*packages:3.0.0-beta(or use the@betadist-tag for other@ai-sdkpackages)
An example upgrade command would be:
pnpm install ai@beta @ai-sdk/react@beta @ai-sdk/openai@betaai package
CoreMessage removed
The deprecated CoreMessage type and related functions have been removed (PR #10710). Replace convertToCoreMessages with convertToModelMessages.
import { convertToCoreMessages, type CoreMessage } from 'ai';
const coreMessages = convertToCoreMessages(messages); // CoreMessage[]import { convertToModelMessages, type ModelMessage } from 'ai';
const modelMessages = convertToModelMessages(messages); // ModelMessage[]generateObject and streamObject deprecated
generateObject and streamObject have been deprecated (PR #10754).
They will be removed in a future version.
You can use generateText and streamText with an output setting instead.
@ai-sdk/openai package
strictJsonSchema defaults to true
The strictJsonSchema setting for JSON outputs and tool calls is enabled by default (PR #10752). This improves stability and ensures valid JSON output that matches your schema.
However, strict mode is stricter about schema requirements. If you receive schema rejection errors, adjust your schema (for example, use null instead of undefined) or disable strict mode.
import { openai } from '@ai-sdk/openai';import { generateObject } from 'ai';import { z } from 'zod';
// strictJsonSchema was false by defaultconst result = await generateObject({ model: openai('gpt-4o'), schema: z.object({ name: z.string(), }), prompt: 'Generate a person',});import { openai } from '@ai-sdk/openai';import { generateObject } from 'ai';import { z } from 'zod';
// strictJsonSchema is true by defaultconst result = await generateObject({ model: openai('gpt-4o'), schema: z.object({ name: z.string(), }), prompt: 'Generate a person',});
// Disable strict mode if neededconst resultNoStrict = await generateObject({ model: openai('gpt-4o'), schema: z.object({ name: z.string(), }), prompt: 'Generate a person', providerOptions: { openai: { strictJsonSchema: false, } satisfies OpenAIResponsesProviderOptions, },});strictJsonSchema no longer used for tool strict mode
Strict mode for tools is now controlled by setting strict on each tool (PR #10817). This enables fine-grained control over strict tool calls, which is important since strict mode depends on the specific tool input schema.
import { openai } from '@ai-sdk/openai';import { streamText, tool } from 'ai';import { z } from 'zod';
// Tool strict mode was controlled by strictJsonSchemaconst result = streamText({ model: openai('gpt-4o'), tools: { calculator: tool({ description: 'A simple calculator', inputSchema: z.object({ expression: z.string(), }), execute: async ({ expression }) => { const result = eval(expression); return { result }; }, }), }, providerOptions: { openai: { strictJsonSchema: true, // Applied to all tools }, },});import { openai } from '@ai-sdk/openai';import { streamText, tool } from 'ai';import { z } from 'zod';
const result = streamText({ model: openai('gpt-4o'), tools: { calculator: tool({ description: 'A simple calculator', inputSchema: z.object({ expression: z.string(), }), execute: async ({ expression }) => { const result = eval(expression); return { result }; }, strict: true, // Control strict mode per tool }), },});structuredOutputs option removed from chat model
The structuredOutputs provider option has been removed from chat models (PR #10752). Use strictJsonSchema instead.
Warning System
Warning Logger Enabled
AI SDK 6 introduces a warning logger that outputs deprecation warnings and best practice recommendations (PR #8343).
To disable warning logging, set the AI_SDK_LOG_WARNINGS environment variable to false:
export AI_SDK_LOG_WARNINGS=falseV2 Model Compatibility Warning
A warning is displayed when using v2 provider models with AISDK v6 (PR #10770). Update your provider implementations to use v3 models.
Tool Changes
Flexible Tool Content Support
AI SDK 6 introduces more flexible tool output and result content support (PR #9605), enabling richer tool interactions and better support for complex tool execution patterns.
Codemods
The AI SDK will provide Codemod transformations to help upgrade your codebase when a feature is deprecated, removed, or otherwise changed.
Codemods are transformations that run on your codebase automatically. They allow you to easily apply many changes without having to manually go through every file.
Codemods are intended as a tool to help you with the upgrade process. They may not cover all of the changes you need to make. You may need to make additional changes manually.
Codemod Table
| Codemod Name | Description |
|---|---|
| TBD | TBD |
Embedding Changes
Embedding Model Method Renamed
The textEmbeddingModel and textEmbedding methods on providers have been renamed to embeddingModel and embedding respectively. Additionally, generics have been removed from EmbeddingModel, embed, and embedMany (PR #10592).
import { openai } from '@ai-sdk/openai';import { embed } from 'ai';
// Using the full method nameconst model = openai.textEmbeddingModel('text-embedding-3-small');
// Using the shorthandconst model = openai.textEmbedding('text-embedding-3-small');
const { embedding } = await embed({ model: openai.textEmbedding('text-embedding-3-small'), value: 'sunny day at the beach',});import { openai } from '@ai-sdk/openai';import { embed } from 'ai';
// Using the full method nameconst model = openai.embeddingModel('text-embedding-3-small');
// Using the shorthandconst model = openai.embedding('text-embedding-3-small');
const { embedding } = await embed({ model: openai.embedding('text-embedding-3-small'), value: 'sunny day at the beach',});Provider Changes
OpenAI
Unrecognized Models Treated as Reasoning Models
The @ai-sdk/openai provider now treats unrecognized model IDs as reasoning models by default (PR #9976). Previously, unrecognized models were treated as non-reasoning models.
This change impacts users who configure @ai-sdk/openai with a custom baseUrl to use non-OpenAI models. Reasoning models exclude certain parameters like temperature, which may cause unexpected behavior if the model does not support reasoning. Consider using @ai-sdk/openai-compatible instead.
Azure
Default Provider Uses Responses API
The @ai-sdk/azure provider now uses the Responses API by default when calling azure() (PR #9868). To use the previous Chat Completions API behavior, use azure.chat() instead.
import { azure } from '@ai-sdk/azure';
// Used Chat Completions APIconst model = azure('gpt-4o');import { azure } from '@ai-sdk/azure';
// Now uses Responses API by defaultconst model = azure('gpt-4o');
// Use azure.chat() for Chat Completions APIconst chatModel = azure.chat('gpt-4o');
// Use azure.responses() explicitly for Responses APIconst responsesModel = azure.responses('gpt-4o');The Responses and Chat Completions APIs have different behavior and defaults.
If you depend on the Chat Completions API, switch your model instance to
azure.chat() and audit your configuration.
Anthropic
Structured Outputs Mode
Anthropic has introduced native structured outputs for Claude Sonnet 4.5 and later models . The @ai-sdk/anthropic provider now includes a structuredOutputMode option to control how structured outputs are generated (PR #10502).
The available modes are:
'outputFormat': Use Anthropic's nativeoutput_formatparameter'jsonTool': Use a special JSON tool to specify the structured output format'auto'(default): Use'outputFormat'when supported by the model, otherwise fall back to'jsonTool'
import { anthropic } from '@ai-sdk/anthropic';import { generateObject } from 'ai';import { z } from 'zod';
const result = await generateObject({ model: anthropic('claude-sonnet-4-5-20250929'), schema: z.object({ name: z.string(), age: z.number(), }), prompt: 'Generate a person', providerOptions: { anthropic: { // Explicitly set the structured output mode (optional) structuredOutputMode: 'outputFormat', } satisfies AnthropicProviderOptions, },});ai/test Changes
Mock Classes Updated
V2 mock classes have been removed from the ai/test module. Use the new V3 mock classes instead for testing.
import { MockEmbeddingModelV2, MockImageModelV2, MockLanguageModelV2, MockProviderV2, MockSpeechModelV2, MockTranscriptionModelV2,} from 'ai/test';import { MockEmbeddingModelV3, MockImageModelV3, MockLanguageModelV3, MockProviderV3, MockSpeechModelV3, MockTranscriptionModelV3,} from 'ai/test';Type Changes
Warning Types Unified
Separate warning types for each generation function have been consolidated into a single Warning type exported from the ai package (PR #10631).
// Separate warning types for each generation functionimport type { CallWarning, ImageModelCallWarning, SpeechWarning, TranscriptionWarning,} from 'ai';// Single Warning type for all generation functionsimport type { Warning } from 'ai';Tool Call Repair Function Signature
The system parameter in the ToolCallRepairFunction type now accepts SystemModelMessage in addition to string (PR #10635). This allows for more flexible system message configuration, including provider-specific options like caching.
import type { ToolCallRepairFunction } from 'ai';
const repairToolCall: ToolCallRepairFunction<MyTools> = async ({ system, // type: string | undefined messages, toolCall, tools, inputSchema, error,}) => { // ...};import type { ToolCallRepairFunction, SystemModelMessage } from 'ai';
const repairToolCall: ToolCallRepairFunction<MyTools> = async ({ system, // type: string | SystemModelMessage | undefined messages, toolCall, tools, inputSchema, error,}) => { // Handle both string and SystemModelMessage const systemText = typeof system === 'string' ? system : system?.content; // ...};