Rohan Taneja

This commit introduces a **new capability** to the **Perplexity language model integration**, enabling it to expose **provider-reported cost information** within the `providerMetadata` of responses. The `doGenerate` and `doStream` methods in `perplexity-language-model.ts` are updated to capture and surface this data directly from the Perplexity API. This enhancement provides users with greater transparency into the operational costs associated with using Perplexity models, significantly improving cost tracking and observability for AI operations. Test snapshots and expectations were updated to reflect the inclusion of this new `cost` field.

This commit **fixes a documentation error** within the **Google provider**'s **multimodal embedding options**. Specifically, it **corrects the JSDoc comment** for the `content` option in `google-generative-ai-embedding-options.ts` to accurately describe how multimodal parts are handled during embedding requests. This **documentation improvement** ensures developers have precise guidance when configuring multimodal content for embeddings, enhancing clarity and preventing potential misconfigurations.

This commit introduces **multimodal embedding support** for the **Google Generative AI provider**, enabling the processing of diverse input types. It's a **new feature** that modifies the `embed` function within `google-generative-ai-embedding-model.ts` to incorporate multimodal content parts for both single and batch embedding requests. A new schema for multimodal content is defined in `google-generative-ai-embedding-options.ts`, adding a `content` option to the embedding model's configuration. This enhancement allows users to leverage Google's capabilities for generating embeddings from inputs combining text, images, and other modalities.

This commit **fixes a bug** within the **Google Generative AI provider's streaming logic** where `groundingMetadata` and `urlContextMetadata` were not correctly preserved if they arrived in earlier stream chunks before the final `finishReason`. The `doStream` function in `google-generative-ai-language-model.ts` was modified to properly store and accumulate this critical contextual metadata throughout the streaming process. New tests have been added to `google-generative-ai-language-model.test.ts` to verify that this information is now consistently available at the end of a streamed response. This **bug fix** ensures the reliability and completeness of metadata associated with streamed AI outputs, improving the accuracy of grounding and URL context for users.

This commit introduces a **new capability** to the **AI Gateway**, allowing users to configure **per-provider timeouts** for AI model requests. It adds a `providerTimeouts` property to the gateway language model options schema, enabling granular control over the duration of `doGenerate` and `doStream` operations for individual providers. This enhancement to the **gateway's core configuration** improves the reliability and responsiveness of AI interactions by preventing indefinite waits. The feature is fully documented in `00-ai-gateway.mdx` and includes comprehensive tests to verify correct timeout propagation.

This commit introduces a **new capability** to the **Prodia image model** by modifying its API requests to include a `price=true` query parameter. This enhancement allows the system to **fetch and extract pricing information** directly from the Prodia API responses. The extracted `dollars` field is then integrated into the image metadata, providing users with transparent cost details for image generation. Test cases for the `ProdiaImageModel` and `ProdiaProvider` were updated to validate this new functionality.

This commit introduces a **new feature** to the **OpenAI image model** within the `openai` package, enabling it to accurately **differentiate between text and image input tokens**. The core logic, implemented in `packages/openai/src/image/openai-image-model.ts`, now distributes these distinct token types evenly across multiple generated images within the provider metadata. This enhancement ensures more precise token accounting and resource allocation for image generation, with new test cases in `openai-image-model.test.ts` verifying the correct distribution behavior.

This commit delivers a **bug fix** for the **`@ai-sdk/gateway`** package, specifically addressing a **schema mismatch warning** encountered during **image generation**. The issue caused unnecessary warnings to be displayed when the Gateway component processed image generation requests, despite the operations being valid. This **patch** resolves the underlying schema discrepancy, ensuring that users of `@ai-sdk/gateway` will no longer see these erroneous warnings, thereby improving the stability and clarity of the image generation process.

This commit introduces a **new feature** to report **image generation usage information** within the **Gateway** service. The `gateway-image-model.ts` module is updated to capture and emit this usage data, and also extends the types of warnings (unsupported, compatibility, mixed) that the image model can return. Comprehensive new test cases in `gateway-image-model.test.ts` validate the accurate reporting of usage and the correct handling of various warning scenarios. This enhancement provides crucial operational insights into image generation activity, supporting better monitoring and potential usage-based analytics.

This commit delivers a **bug fix** for the **`@ai-sdk/openai-compatible` package**, specifically addressing the omission of `reasoning_content` in assistant messages during multi-turn tool calls. It modifies the `convertToOpenAICompatibleChatMessages` function to correctly extract and include this content from assistant message parts. An optional `reasoning_content` field was also added to the `OpenAICompatibleAssistantMessage` interface to support this data. This ensures **improved data fidelity and compatibility** for systems utilizing the OpenAI-compatible adapter, preventing the loss of important internal reasoning information.

This commit introduces a **bug fix** for the **TogetherAI image generation model**, specifically addressing how the `n` parameter is handled during image requests. The `doGenerate` method in `packages/togetherai/src/togetherai-image-model.ts` has been updated to **conditionally include the `n` parameter** in the request payload only when its value is greater than 1, preventing redundant transmission of `n: 1` for single image requests. This change optimizes API calls to TogetherAI by ensuring parameters are sent only when necessary. Accompanying **test updates** in `togetherai-image-model.test.ts` validate this behavior, removing superfluous `n: 1` checks and adding a dedicated test for multiple image generation requests.

This commit **fixes** an issue within the **`openai-compatible` package** where the image generation models incorrectly hardcoded `providerOptions` to only process options under the `openai` key, regardless of the actual provider. It **improves developer experience** by enabling dynamic extraction of provider-specific options, allowing users to correctly specify configuration using the actual provider's name (e.g., `recraft`) as the key in `providerOptions`. This change enhances flexibility for configuring diverse **OpenAI-compatible providers**, ensuring that custom settings are applied as intended. The **`openai-compatible-image-model.ts`** module is updated, alongside new test cases and revised documentation examples to reflect this improved configuration.

This commit **updates the model settings files** within the **`gateway` provider** to **expand support for a diverse range of AI models**. It introduces new **embedding models** from Alibaba and Voyage, **image models** from Google, and numerous **language models** from providers including Amazon, Mistral, OpenAI, and others. This **feature enhancement** ensures the `gateway` service offers the latest AI capabilities, directly impacting users by providing a broader selection of models and updating `autocomplete` functionality.

This commit introduces a **new capability** by adding a dedicated **`@ai-sdk/prodia` provider package** to the AI SDK. This package enables seamless integration with **Prodia's image generation API**, which is not OpenAI-compatible, allowing users to leverage Prodia's services directly. It includes the implementation of `ProdiaImageModel` for handling image generation requests and responses, along with comprehensive unit tests for the new provider. A new example script, `examples/ai-core/src/generate-image/prodia.ts`, is also added to demonstrate its usage. This significantly expands the range of supported AI providers for image generation within the SDK.

This commit provides a **bug fix** for the **`google` provider package** that resolves an issue with **tool schema conversion** when interacting with Google Gemini models. Previously, the `convertJSONSchemaToOpenAPISchema` utility incorrectly converted nested empty object properties (e.g., `{ type: "object" }`) to `undefined`, causing validation errors because the `required` array still referenced them. The fix modifies this function to **preserve nested empty objects** as `{ type: "object" }` while still converting root-level empty objects, ensuring **correct tool schema validation** and preventing runtime failures. This change specifically impacts how **tool definitions** are processed before being sent to Gemini, preventing "property is not defined" errors. New examples and tests have been added to verify this behavior.

This commit introduces support for the new `gpt-5.1-codex-max` model within the **OpenAI provider**. It **adds** this model ID to the internal list of recognized OpenAI response and reasoning models in `packages/openai/src/responses/openai-responses-options.ts`. This **feature enhancement** allows applications using the **OpenAI provider** to leverage the capabilities of this new model, expanding the range of available options for AI interactions.

This commit introduces **support for the `xhigh` reasoning effort** option within the **OpenAI provider** of the AI SDK. It **extends the `reasoningEffort` enum** in `packages/openai/src/chat/openai-chat-options.ts` to allow specifying this new level for both chat and response language models. This **new feature** enables users to request higher reasoning capabilities from compatible OpenAI models. Comprehensive **documentation updates** in `content/providers/03-openai.mdx` and `packages/openai/src/responses/openai-responses-options.ts` detail its usage and model availability, alongside **new test cases** ensuring correct implementation.

This commit introduces **support for Nova 2 extended reasoning** within the **Amazon Bedrock provider**, enabling the use of the `maxReasoningEffort` field for enhanced model capabilities. It **adds a new capability** to the `bedrock-chat-options.ts` schema and implements the necessary logic in `bedrock-chat-language-model.ts` to process this configuration. The change includes validation and warnings for incompatible reasoning configurations across different Bedrock models, such as Anthropic, ensuring robust integration. Documentation in `amazon-bedrock.mdx` and comprehensive test cases in `bedrock-chat-language-model.test.ts` are updated to reflect this new functionality and its expected behavior.

This commit introduces **new provider options** for the **Black Forest Labs image model**, significantly enhancing its configurability. It implements support for explicit image `width` and `height`, `steps`, `guidance`, and the ability to process multiple input images within the `packages/black-forest-labs` module. This **new feature** provides users with more granular control over image generation parameters, improving the flexibility and utility of the Black Forest Labs provider. The changes include updates to the `blackForestLabsImageProviderOptionsSchema`, corresponding **documentation**, and **test cases**.

This commit introduces a **new capability** within the **`gateway`** package, enabling **server-side image request splitting**. It modifies `packages/gateway/src/gateway-image-model.ts` by changing the `maxImagesPerCall` property from a dynamic getter to a large static value. This crucial change **prevents client-side image request splitting**, ensuring that image requests are batched and processed more efficiently by the server. The update includes corresponding test adjustments in `gateway-image-model.test.ts` and documentation via a new changeset, streamlining the overall image request handling workflow.