@ai-sdk/openai

@ai-sdk/provider-utils is responsible for the general-purpose utility functions in the AI SDK for interacting with model providers, including low-level logic such as streaming response parsing, tool call formatting, and error handling.

The upgrade method is to run npm install @ai-sdk/openai@4.0.60.

The AI SDK adopts a Provider architecture in which each Provider package implements standardized interfaces such as LanguageModelV1, which is an engineering practice of the strategy pattern.

Within the AI SDK ecosystem, there exists an adapter package @ai-sdk/zod similar to @ai-sdk/valibot, with both providing the same public interface.

The Vercel AI SDK's unified multi-vendor abstraction layer is implemented based on a Provider plugin mechanism, where each vendor adapter is responsible for converting that vendor's raw HTTP API into the SDK's internal standard LanguageModelV1 interface.

@ai-sdk/provider defines unified TypeScript interfaces such as LanguageModelV1 and EmbeddingModelV1, allowing upper-layer applications to remain unaware of the underlying vendor implementation details.

The @ai-sdk/openai package is responsible for integration with OpenAI models.

The fact that @ai-sdk/amazon-bedrock depends on @ai-sdk/openai does not mean an OpenAI account or key is required, because @ai-sdk/openai encapsulates general-purpose capabilities such as HTTP request handling, streaming response parsing, and tool-calling protocols that are shared across multiple provider packages.

Patch versions are typically backward compatible and do not introduce breaking changes.

@ai-sdk/openai@3.0.121 was released as a patch version, with the core change being the addition of support for the GPT-6.1 Sol model.

The AI SDK includes built-in adapters for multiple model providers such as OpenAI, Anthropic, and Google, so developers only need to switch the provider to change the underlying model.

The ai package is a high-level wrapper for the application layer, exposing APIs such as generateText and streamText that developers call directly.

All concrete provider packages (such as @ai-sdk/openai, @ai-sdk/anthropic, @ai-sdk/xai) must implement the interfaces declared by @ai-sdk/provider, and any enum extension, field addition or removal, or type narrowing in that package propagates to all downstream providers

The core fix in this update is the rewriting of recursive Zod schemas that use allOf wrappers, with commit record 4e94782.

This fix detects and unwraps redundant allOf wrappers in the conversion layer, making the generated JSON Schema compliant with OpenAI's strict mode requirements.

The article recommends that when upgrading @ai-sdk/azure to 4.0.82, users should simultaneously confirm that @ai-sdk/openai is at version 4.0.78 or a compatible version, and verify Azure OpenAI calls in a test environment.

The strategy of directly rejecting at the history level while issuing warnings and ignoring at the request level balances runtime robustness with data reliability and is a deliberate trade-off design.

For unsupported effort updates in historical messages, the SDK rejects them directly.

The SDK validates update requests against the effort levels actually supported by the model, issues warnings for unsupported updates at the request level, and automatically ignores them.

@ai-sdk/azure reuses a large amount of the underlying logic from @ai-sdk/openai, with differentiated handling only at the levels of authentication method, endpoint configuration, and deployment name.