Vercel AI Updates: July 2, 2026

1. Vercel Adds Service Bindings for Secure Internal Communication Between Services

Vercel. Vercel introduced Service Bindings, which let one service securely call another within the same deployment without exposing it publicly. When a service declares a binding, Vercel injects an internal URL through an environment variable so the caller can make standard fetch() requests, with TLS, routing, and authentication handled automatically over Vercel’s internal network. A service stays private unless explicitly exposed, and service-to-service calls appear in observability and are billed as Service Requests and Fast Origin Transfer rather than CDN requests. Source

2. Claude Fable 5 Access Restored on Vercel AI Gateway

Vercel. Vercel reinstated access to Anthropic’s Claude Fable 5 on its AI Gateway after the US government lifted export restrictions, with the model unchanged from its earlier June availability but shipping updated, more robust safety classifiers. Vercel warns that the enhanced safety filters may occasionally block routine tasks such as coding work and recommends configuring model fallbacks so requests still complete if Fable 5 declines them. Prompts and completions for the model are retained for 30 days without Zero Data Retention support, as cumulative request patterns are used to detect misuse. Source

3. AI SDK Vertex Provider Adds Gemini Interactions API for Multimodal Output

Vercel. Vercel shipped @ai-sdk/google-vertex@5.0.8, adding a vertex.interactions() method for the Gemini Interactions API on Vertex AI that targets the location-scoped interactions resource. The addition enables multimodal-output models such as gemini-omni-flash-preview, which supports video output, through Vertex using existing OAuth credentials. The release also exports the GoogleInteractionsLanguageModel from @ai-sdk/google/internal for provider reuse. Source