AI SDK 5: What’s New and What’s Next

As models gained reasoning, tools, and multimodal input and output, applications became more complex too. Incremental changes made provider abstractions harder to maintain: Anthropic’s signed reasoning and OpenAI’s newer reasoning APIs were difficult to fit into the existing specification. The team started the three-month AI SDK 5 project in April.

AI SDK 5 moves the provider specification to version two, with more room for provider-specific information. Grammel describes ordered multimodal outputs, provider-executed tools such as OpenAI web search, and preserving response IDs so longer conversations work with APIs such as OpenAI Responses. The SDK also adds transcription and speech generation APIs.

The chat redesign lets applications define typed message metadata, data parts, and tool calls across both server and front end. A travel app, for example, could stream typed hotel bookings or activities and render them with custom components. Chat is independent of a particular UI framework, and developers can replace the default Server-Sent Events transport or use `streamText` directly in a React Native app.

The SDK adds controls for agentic tool-calling loops without requiring a full agent framework. With `prepareStep`, developers can adjust the model, prompt, messages, or temperature for each step. With `stopWhen`, they can set limits such as a maximum number of steps or implement custom conditions based on tool calls or spending.

Grammel says version 5.1 is expected around October, with improved telemetry to help developers observe model calls and support evaluation loops. The team is also working on image editing, which has multiple use cases such as selecting and refining an area. Streaming speech generation and other capabilities may come later.