The latest AI news and updates of Vercel — Frontend cloud platform and creator of the AI SDK for building AI-powered web applications. Covering Vercel's latest product updates, launches, and analysis from the past 90 days.
v0Vercel8h agoVercel now supports WebSocket connections for Python Functions, enabling bidirectional communication for real-time features like AI streaming and live collaboration. The update covers both ASGI and WSGI applications, including support for FastAPI, Django, and Flask. This allows persistent connections between clients and server-side code on the platform.
VercelJul 28Vercel added new configuration options for Kimi K3 on AI Gateway, including a fast speed tier, US-based inference, and zero data retention. These settings optimize performance and compliance for the model without additional token markups. The updates are available for integration through the Vercel AI SDK.
VercelJul 27Vercel released DeepsecBench, a benchmark evaluating how effectively AI models identify cybersecurity vulnerabilities in application code. GPT-5.6 Sol achieved the highest score, while Kimi K3 provided half the performance at one-fifth the cost. Grok 4.5 delivered the best score-to-cost ratio among the top ten models. The benchmark uses a weighted F2 score to prioritize recall over precision.
VercelJul 27Vercel added regional inference to AI Gateway, allowing requests to be pinned to US or EU data centers. The new inferenceRegion field ensures data residency compliance by failing closed if the requested region is unavailable. Every response confirms the specific region that served the request, providing an auditable trail for production AI applications.
VercelJul 27Vercel added WebSocket mode to its AI Gateway for the OpenAI Responses API, enabling persistent connections across conversation turns. This reduces end-to-end latency by approximately 40% for agentic workflows with 20 or more tool calls. The mode supports Zero Data Retention and store=false configurations, requiring only new inputs and response IDs per turn.
VercelJul 25Vercel Labs introduced a built-in terminal component for its Native SDK, providing a real shell on a PTY powered by libghostty-vt. The component supports text selection, scrollback, and truecolor, while enabling session recording and offline replay. It is available for macOS, Windows, and Linux to support software factories and agent orchestrators.
v0Jul 24v0 now converts full Figma files into working applications from a single link. The agent autonomously explores pages and frames, extracting design tokens, layout, and Dev Mode assets to construct the UI. It validates each generated screen against the original frame image during the build process to ensure visual fidelity.
VercelJul 23Vercel added streaming transcription to AI Gateway, enabling the processing of live audio streams with real-time transcript updates. The new streamTranscribe function in the AI SDK handles raw audio chunks as they arrive, reducing latency for voice agents. This beta feature supports any streaming-capable transcription model, including openai/gpt-realtime-whisper and xai/grok-stt.
VercelJul 23Vercel added Ant Group's Ling 3.0 Flash to AI Gateway, a 124B-parameter Mixture-of-Experts model featuring a 256K token context window. The model supports both thinking and non-thinking modes for token-efficient agentic inference. It is available for free through August 3rd with no platform markup on inference costs.
VercelJul 22Vercel added service tiers to AI Gateway, enabling per-request control over latency and cost. The priority tier increases token throughput at 1.8-2x the default rate, while the flex tier reduces costs to 0.5x for latency-tolerant tasks. Billing adjusts automatically based on the tier served, with support for OpenAI and Gemini models across all Gateway API formats.
VercelJul 19Vercel Labs updated agent-browser with native HAR network recording and a new derive-client skill. The CLI captures network traffic during browser sessions, allowing agents to generate programmatic API clients. This transition from full browser control to direct API interaction reduces overhead for multi-step agentic tasks.
VercelJul 19Vercel released the AI SDK for Python in public beta, bringing agent loops, tool calling, and streaming to the Python ecosystem. The toolkit includes built-in primitives for managing message history, asynchronous tool dispatch, and multi-model support. Installation is available via uv add ai.
VercelJul 17Vercel launched its platform plugin for the Kimi Code CLI, providing AI agents with on-demand knowledge of Next.js, the AI SDK, and Vercel Functions. The plugin ensures agents use current platform APIs and recommended patterns during development. Installation requires upgrading to the latest Kimi Code CLI and selecting the plugin from the third-party menu.
VercelJul 17Vercel provides a 35% discount on GLM 5.2 model usage when routed through Novita on AI Gateway. This promotion remains active through July 24, 2026. After the discount period ends, the model continues to be available at standard provider rates with no platform markup.
VercelJul 16Vercel launched the Vercel Plugin for GitHub Copilot CLI and VS Code, adding 30 platform skills, three specialized agents, and four built-in commands. The plugin provides real-time code validation for Next.js, the AI SDK, and Vercel Functions, ensuring coding agents utilize current platform APIs and recommended patterns during development.
VercelJul 16Vercel integrated Moonshot AI's Kimi K3 model into its AI Gateway, providing access to the open-source model via the AI SDK. Kimi K3 features a 1-million-token context window and native visual understanding for text, image, and video inputs. The integration supports Bring Your Own Key authentication and reflects provider pricing with no platform markup or inference fees.
VercelJul 16Vercel released the @agent-browser/eve extension, which mounts browser automation capabilities directly into eve agents. The package provides namespaced tools for navigating websites, filling forms, and capturing screenshots within a sandboxed environment. These tools operate independently of the main application runtime, allowing agents to perform web-based tasks and data collection autonomously.
VercelJul 15Vercel added Thinking Machines' Inkling to AI Gateway, making the lab's first open-weights multimodal model available via the AI SDK. The model uses a Mixture-of-Experts architecture with a 1-million-token context window and controllable thinking effort. It supports agentic, reasoning, coding, vision, and audio tasks, accessible through the thinkingmachines/inkling identifier with no platform markup on inference.
VercelJul 15Vercel integrated ByteDance’s Seedream 5.0 Pro image generation model into its AI Gateway. The model renders text and infographics alongside realistic imagery, accessible via the AI SDK. The integration supports Bring Your Own Key authentication and reflects provider pricing with no platform markup or inference fees.
VercelJul 15Vercel has opened its AI Gateway leaderboard data under a CC BY 4.0 license. The dataset, updated daily, tracks real production usage across models, labs, apps, and providers, including metrics like token volume and spend. The data is downloadable as CSVs, queryable via a new API endpoint, and exportable as shareable chart images.
VercelJul 15Vercel integrated AgentMail into its Marketplace, providing AI agents with dedicated email inboxes. The integration handles mail servers and deliverability while supporting full thread memory and structured data extraction. Agents trigger workflows via webhooks and WebSockets, with provisioning available directly through the Vercel dashboard or the Vercel CLI.
VercelJul 13Vercel’s July AI Gateway Production Index reports open-weight models reached 29% of token volume in June, up from 11% in April, while total spend remained flat. DeepSeek captured 22.6% of volume, while Anthropic maintained 61% of spend. Additionally, US export controls briefly suspended Claude Fable 5, which had reached 22% of Opus 4.8’s request volume within four days of launch.
VercelJul 13Vercel Labs released Native SDK v0.5, adding TypeScript authoring for app cores. The toolkit compiles TypeScript into arena-backed native code with 83ns dispatch, removing the need for a JS engine or garbage collector. This update reduces agentic coding overhead, cutting turns, cost, and lines of code for native desktop applications.
VercelJul 10Vercel now supports Lovable applications with zero configuration. Syncing a Lovable project to GitHub and importing the repository into Vercel triggers automatic framework detection for TanStack Start. Since Lovable projects utilize the Nitro server toolkit, every subsequent change in Lovable automatically syncs to GitHub and initiates a new deployment on Vercel.
VercelJul 9Vercel added Meta’s Muse Spark 1.1 to its AI Gateway, providing access to the multimodal reasoning model for agentic tasks. The integration supports the meta/muse-spark-1.1 model identifier with Bring Your Own Key authentication and no platform markup or inference fees. The model features a 1-million-token context window and accepts text, image, video, PDF, and audio inputs.
VercelJul 8Vercel added SpaceXAI’s Grok 4.5 model to its AI Gateway, supporting text and image inputs with configurable reasoning levels. The integration includes no-markup inference and Bring Your Own Key support. A single vercel ai-gateway rules add command redirects existing Grok 4.3 traffic to the new model.
VercelJul 8Vercel launched GitHub tools for its eve agent framework, creating GitHub agents in nine lines of code. A single file registers tools using presets like maintainer or code-review. Risky write actions require human approval by default, and read tools automatically trim payloads to manage context window usage.
VercelJul 5Vercel now provides Agent Run observability for the eve framework through its MCP server and CLI. Four new tools and commands allow inspection of trace data, including reasoning, tool calls, and token usage. These traces are automatically ingested upon deployment, providing direct access to metadata and lifecycle events for debugging and agent improvement without leaving the editor.
VercelJul 2Vercel released ai-cli v0.4.2, introducing the ai models command for terminal-based model inspection. Passing a model ID now displays its capabilities, context window, pricing, and provider-specific metrics like latency and uptime. The update also includes a --json flag for programmatic access to this metadata, allowing model details to be integrated directly into scripts.
v0Jul 1v0 has restored access to Anthropic's Claude Fable 5 model within its AI web application generation platform. The model, which handles complex, long-running tasks, is now available again for selection directly from the v0 prompt bar.
VercelJul 1Vercel has open-sourced konsistent, a CLI linter that enforces structural conventions in TypeScript codebases. The tool, which Vercel uses internally for its AI SDK and Chat SDK, ensures that both human developers and AI agents maintain consistent API structures. The konsistent skill helps agents generate project-specific configuration files to automate these rules.
VercelJul 1Vercel launched a suite of agentic infrastructure tools at Ship 2026, including the eve agent framework, Vercel Services for full-stack microservices, and Dockerfile support. The platform adds an autonomous dashboard agent, AI SDK 7 with agent harnesses, and enterprise security features like Vercel Passport and Connect. These updates provide a unified environment for building, deploying, and securing production-grade AI agents.
VercelJul 1Vercel updated its AI CLI to version 0.4.0, adding native voice capabilities via the AI Gateway. The new ai audio speak command generates speech from text or stdin, while ai audio transcribe converts audio files to text. Both commands support model selection, file output, and inline terminal playback or waveform visualization for compatible environments.
VercelJun 30Vercel and Shopify are rebuilding Hydrogen from the ground up to be agent-first and runtime-agnostic. The framework now runs anywhere JavaScript does, removing dependencies on specific runtimes or proprietary glue code. A developer preview for Next.js is available now, with support for additional frameworks planned for future releases.
VercelJun 30Vercel launched its Agent into public beta, adding dashboard chat, production investigations, and user-approved remediation actions. The agent traces deployment issues and proposes fixes, including PRs or rollbacks, which require explicit sign-off before execution. Simple queries remain free, while investigative and code-writing tasks incur a $0.25 per million token rate for Pro and Enterprise teams.
VercelJun 28Vercel Labs released agent-browser 0.31, introducing durable session memory for AI browser agents. The update enables worktree-scoped sessions that automatically restore state, perform validation checks, and execute transactional saves. These features allow agents to maintain continuity across browser tasks, improving reliability for multi-step agentic workflows.
VercelJun 28Vercel added adapters for LangChain Deep Agents and OpenCode to the AI SDK Harness API. The HarnessAgent interface runs these coding-agent runtimes inside a Vercel Sandbox, supporting runtime swaps without changing application code. This update brings the total supported harnesses to five, including Claude Code, Codex, and Pi.
VercelJun 25Vercel released a product-design system that teaches coding agents to follow internal design standards. The system integrates a repository-based skill for context, deterministic linters for rule enforcement, and evals to test guidance. A weekly evidence-intake loop collects design feedback from Slack and GitHub, providing a structured process for human review and updates to rules and examples.
v0Jun 25v0 Design Systems 2.0 imports design systems from GitHub, npm, Storybook, Figma, and real apps. The platform learns components, tokens, and global styles, saving them as a reusable skill. When attached to a chat, v0 generates UI using the team's actual production code and conventions. Private packages are supported through shared environment variables.
VercelJun 23Vercel’s eve framework introduces a built-in evaluation system that tests agents against real runtimes. The system uses defineEval to verify tool usage, dangerous tool avoidance, and response accuracy. It provides three assertion surfaces—run-level methods, value checks, and LLM-as-judge grading—to catch regressions without requiring a separate, simulated testing environment.
VercelJun 23Vercel is now a one-click deployment destination within Anthropic's Claude Design. The integration deploys designs as new projects and returns a live URL directly from the Share menu. Connecting the Vercel MCP server enables this direct handoff, moving designs from the canvas to production hosting without leaving the interface.
VercelJun 22Vercel integrated Sakana AI's Fugu Ultra into its AI Gateway, providing Mythos-class intelligence through multi-agent orchestration. The model coordinates a pool of frontier models to route tasks to specialized agents, delivering results in a single call. It is available via the AI SDK with no platform markup and supports Bring Your Own Key authentication.
VercelJun 21Vercel enables Z.ai's GLM-5.2 model for use within Anthropic's Claude Code via the Vercel AI Gateway. The integration routes requests through the gateway's Anthropic-compatible endpoints, providing observability and traffic monitoring for the open-weights model. Configuration of environment variables routes Claude Code sessions through the gateway, enabling the model's high-effort reasoning mode for agentic coding tasks.
VercelJun 17Vercel is introducing a platform for enterprise apps and agents that integrates identity, access control, and audit trails by default. The bundle includes Vercel Passport for identity provider authentication, Vercel Connect for scoped, short-lived agent credentials, Enterprise Managed Users for directory-synced lifecycle control, and BYOC on AWS for private infrastructure. Passport and Connect are currently in public beta.
VercelJun 17Vercel Connect entered public beta, providing a secure way for AI agents to access external services like Slack, GitHub, and Salesforce. The platform issues short-lived, task-scoped tokens at runtime, eliminating the need to store long-lived provider secrets in environment variables. Connectors support custom APIs and MCP servers, featuring built-in webhook verification and centralized token revocation.
VercelJun 17Vercel launched eve, an open-source framework for building and scaling AI agents using a filesystem-based directory structure. The framework includes built-in durable execution, sandboxed compute, human-in-the-loop approvals, and tracing. Agents deploy to Vercel with one command and support multiple communication channels, including Slack and Discord, for production use.
VercelJun 16Vercel has integrated Z.ai’s GLM-5.2 model into its AI Gateway, providing developers with a 1-million-token context window for long-horizon engineering tasks. The model ranks fifth on Next.js evaluations, achieving an 88% base score and 96% with agentic instructions. It is accessible via the AI SDK with no platform markup or provider account requirements.
VercelJun 16Vercel Labs launched agent-browser as an MCP server, adding a second integration path for the browser automation tool alongside its existing setup. The Model Context Protocol wrapper suits developers building from MCP-compatible environments like Cursor, or those who want finer control over how their agent accesses the browser. The underlying automation engine and behavior are unchanged.
VercelJun 16Vercel added Auth0 as a one-click integration to its Marketplace, providing production-ready authentication for AI applications and agents. The integration automatically provisions Auth0 applications, supports the Auth0 Next.js SDK, and synchronizes authentication configurations across Development, Preview, and Production environments. This update removes manual setup for identity management in agentic workflows.
VercelJun 16Vercel now enables Claude Code, Codex, and Pi agents to render interactive user interfaces like charts, forms, and 3D components directly within sandboxed sessions. This integration uses the AI SDK's experimental HarnessAgent and the json-render framework to transform raw AI output into structured, visual dashboards during task execution.
v0Jun 16v0 now supports skills, which the AI agent applies across every generation. Skills attach directly from the prompt bar, pulling from the skills.sh directory, saved collections, or local repositories. The agent uses these packaged instructions to maintain consistent behavior and domain-specific capabilities throughout the web application generation process.
VercelJun 13Vercel launched HarnessAgent in AI SDK v7, a unified API for running agent harnesses like Claude Code, Codex, and Pi. The abstraction normalizes access to sandboxed sessions and agent runtimes, making harnesses swappable without rewriting application code. HarnessAgent streams are compatible with existing AI SDK tooling, including UI hooks like useChat.
v0Jun 13v0 introduced annotations, enabling direct element selection in live previews to attach comments. These comments bundle into a single prompt for the AI agent to execute design changes. This update provides a visual feedback loop for iterating on web interfaces, replacing the need for verbose text-based instructions.
VercelJun 13Vercel launched its plugin for Grok Build, enabling AI coding agents to autonomously deploy and manage web projects. The plugin dynamically injects Vercel-specific knowledge into the agent's context, ensuring code stays aligned with platform APIs and recommended patterns. Developers can install the plugin via the Grok Build CLI marketplace or through prompt-based recommendations.
VercelJun 13Vercel added Azure as a provider for DeepSeek V4 Pro and V4 Flash on its AI Gateway. Requests now automatically route through Azure for improved failover, or Azure can be prioritized via the order option in the AI SDK. The integration supports Bring Your Own Key authentication with no platform fees or inference markups.
VercelJun 13Vercel Labs released agent-browser v0.27.2, turning silent failures into actionable errors. Off-screen clicks now scroll into view, dialogs no longer hang, and blocked clicks identify the covering element. Additionally, the update cuts warm CLI command latency from 150ms to 1ms, improving responsiveness for agentic workflows.
VercelJun 13Vercel AI Gateway now lets API keys carry a dollar spending budget. Once a key hits its cap, the gateway rejects further requests until the budget resets or is raised. Each budget takes an optional daily, weekly, or monthly refresh period, and is set in the Vercel dashboard or programmatically via the Vercel CLI.
v0Jun 10Vercel's v0 has integrated Anthropic's Claude Fable 5 model. This makes a frontier AI model, known for handling complex, long-running tasks, available within an AI-powered web application builder for Premium and Team plan users.
VercelJun 10Vercel has made Anthropic's Claude Fable 5, a Mythos-class model, available on its AI Gateway. This integration provides developers with access to a highly capable AI for complex, multi-step tasks without requiring separate provider accounts.
VercelJun 5Vercel Sandbox now supports persistent drives in private beta. This separates storage from compute, allowing AI agents to retain their workspaces and "memories" across disposable sandbox instances. This is crucial for building reliable, multi-step agentic systems that need to maintain state.
Frequently asked questions