BreakpointRogo uses Vercel to deploy agent-written code to production in about five minutes, reaching more than 73,000 deployments per month. The company also uses AI SDK agent swarms to triage and remediate production incidents without manual intervention.
Vercel engineering blog
Vercel introduces Jev, a fast structured-decision model, and shows Python engineers how to use it through the AI SDK's experimental evaluate() API. The post explains Jev's classifier-oriented design and evaluates it for Python-versus-English detection and AST-based Python code generation.
Vercel will stop recording new First Input Delay (FID) measurements in Speed Insights on November 1, 2026. Existing FID data will remain available, while Interaction to Next Paint (INP) continues as the responsiveness metric in Real Experience Score with no user action required.
Vercel announces that Convai Innovations’ Laya evaluation model is available through AI Gateway and free via Boundless until October 31, 2026. The post demonstrates using Laya through the AI SDK evaluation API for classification, routing, guardrails, and rubric-based scoring.
Vercel AI Gateway now supports Microsoft’s MAI Voice 2.1, Voice 2.1 Flash, and Transcribe 2 Streaming models for multilingual speech generation and live transcription. The post demonstrates using AI SDK functions to generate audio and process partial transcript updates from streaming PCM input.
Vercel’s AI Gateway now supports Browserbase Search and Fetch tools, enabling models to discover current web information and retrieve page contents across providers. The update covers the required AI SDK version, API key configuration, and tool setup for search and fetching.
Vercel has renamed the Edge Requests usage metric to CDN Requests without changing pricing, limits, or measurement. Integrations should continue using stable SkuId values rather than matching the renamed ServiceName field in billing charges API responses.
Vercel’s AI Gateway now supports InclusionAI’s Ling 3.1 Flash, a 560B-parameter hybrid reasoning model with 25B active parameters and a 262K-token context window. Developers can call it through a unified API for coding, long-document analysis, tool-using agents, usage tracking, routing, retries, and failover.
Vercel Agent can now install private npm packages and dependencies from custom registries during agent sessions. Teams authenticate by configuring shared NPM_TOKEN or NPM_RC environment variables, with credentials kept outside the agent sandbox.
Vercel Sandbox can now connect to a team’s dedicated Secure Compute network, providing static outbound IPs and access to private AWS resources through VPC peering. The update supports network attachment through the JavaScript API or CLI, with changes applied to subsequent sessions.
Vercel’s CDN no longer stores origin responses whose Vary header includes Cookie, though it continues serving them normally. The changelog explains how to identify these responses with x-vercel-cache and Runtime Logs, and recommends removing Cookie from Vary for non-personalized content or using Cache-Control: private for personalized responses.
Vercel Connect now allows users to submit services to its connectors directory. The workflow supports OAuth and API key authentication, requires verification with a real token call, and sends submissions through Vercel review.
Vercel CLI now supports searching project trace spans directly from the terminal. Developers can filter by status, service, environment, identifiers, and time, run supported KQL queries, and export span results as JSON Lines for scripts and coding agents.
Vercel announces that OpenAI's GPT-6.1 Sol is available through AI Gateway for coding, computer-use, document analysis, and agent workflows. The post shows how to call the model with the AI SDK or compatible APIs and configure it for coding agents using the Vercel CLI.
Vercel now allows users to search for domain suggestions, availability, and pricing without authentication. The update covers CLI keyword searches and unauthenticated Domains Registrar API requests for checking up to 200 exact domains, while purchases and management still require authentication.
Vercel makes Anthropic’s Claude Sonnet 5.5 available through AI Gateway under the model identifier anthropic/claude-sonnet-5.5. The post shows how to invoke it with the AI SDK and notes compatibility with OpenAI and Anthropic APIs, coding agents, routing, retries, failover, and Zero Data Retention.
Vercel AI Gateway now supports Fireworks’ Ember-1, a research-preview reasoning model for coding and agentic workflows. The post highlights its reduced reasoning-token usage, 1M-token context window, multimodal and tool-calling support, and provides API and CLI setup examples.
Vercel analyzes seven months of skills.sh data, covering one million agent skills and nearly 280 million installs. The report examines demand across work categories, concentration of usage, cross-industry portability, and the shift toward company-specific skills measured by effectiveness.
Vercel’s Marketplace database browser now supports Redis and Upstash for Redis integrations directly in the dashboard. Users can run commands, execute MULTI/EXEC transactions, browse and filter keys, and inspect value metadata such as type, TTL, encoding, and size.
Vercel announces Pixel Canary, a stealth model available free for a limited time through AI Gateway. The post covers its coding and frontend strengths, Next.js evaluation results, API usage with the Vercel AI SDK, and CLI setup for coding agents.