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, company news, and launches from the past 90 days.
v0
Vercel6h agoVercel added 14 ElevenLabs audio models to its AI Gateway, including tools for realtime voice agents, music, sound effects, and multi-voice dialogue. The integration supports streaming speech over WebSockets and transcription through a single API key. All ElevenLabs models on the gateway are 50% off through November 9, with no platform markup on inference requests.
v06h agov0 now generates self-contained artifacts including reports, charts, diagrams, and interactive HTML pages directly within the chat interface. These outputs are created and edited through natural language prompts, stored outside the project directory, and can be published to share. The feature supports eight file formats, including JSON, CSV, and PDF, with sandboxed execution for interactive HTML.
Vercel6h agoVercel updated HarnessAgent to support image inputs across all ten harness adapters. The AI SDK now accepts images alongside text in the standard message format, with each harness's underlying vision model processing the file. This update applies to adapters including Claude Code, Cursor, GitHub Copilot, and Codex, where agents process visual content directly within the existing harness workflow.
VercelOct 26Vercel Sandbox now supports Scale AI’s AgentEnv, a framework for testing and scoring AI agents. The Vercel provider, available in AgentEnv 0.9.1334 and later, executes the agent and its task environment in separate sandboxes. The Vercel CLI command ./vercel-eval executes evaluations, providing access to agent actions and task scores.
VercelOct 26Vercel added the discover command to its CLI, providing search access to the 1000+ services in the Vercel Connect catalog by name, alias, keyword, or description. The command supports filtering by category, connection method, or webhook triggers, with JSON output for automated workflows. Discovered service names integrate directly into the create command for connector setup.
VercelOct 26Guillermo Rauch built cpu-chat, a tool demonstrating that agents in isolated Firecracker microVMs can communicate covertly via CPU contention, cache timing, and DRAM bandwidth. The reproduction works on x86 and M5 chips, bypassing Firecracker production recommendations. Rauch noted that while proprietary models refused the task, open-weight models successfully performed the analysis, which he argues is essential for building secure sandboxes.
VercelOct 24Vercel has deployed practical mitigations and always-on monitoring for cache-timing attacks within its sandbox environments. These measures address noisy neighbor risks where agents on shared host machines detect cache noise from co-tenants. This class of vulnerability, known since 2015, involves agents communicating through CPU cache contention rather than network or file systems.
VercelOct 23Vercel added Moonshot AI's Kimi K3.1 Pro and a 2–3× faster Highspeed variant to AI Gateway. The 1M-token context model improves performance in coding, 3D modeling, and financial analysis. Access is available via the AI SDK, standard chat APIs, or connected coding agents, with open weights scheduled for release by November 4, 2026.
VercelOct 23Vercel Agent now uses a plan-based permission model to safely investigate and fix production issues. The agent performs read-only investigations by default, requiring explicit user approval for any plan requesting elevated scopes. These plans define specific tasks and access requirements, ensuring the agent operates only with credentials matching the user's own permissions and the approved work.
VercelOct 22Vercel added Baseten's Mammoth-1 and Elephant-1 decision models to AI Gateway, free through October 27. These models classify text, route requests, and score content against defined criteria, returning structured probabilities and choices. Mammoth-1 ranks first on the Jev Decision Index, while Elephant-1 is tied for first on the Artificial Analysis Decision Index.
VercelOct 21Vercel added DeepSeek V4.1 Flash to Token Pass, featuring new 1017 weights and a 1-million-token context window. Token Pass is a $30 monthly subscription providing $80 in AI Gateway credit, now supporting DeepSeek alongside GLM 5.3 and Kimi K3. Existing Token Pass configurations automatically include the new model without requiring changes.
v0Oct 21v0 now supports the Slack MCP, providing access to Slack threads and files for project context. The integration posts progress updates directly back to Slack channels. v0 visibility remains limited to the specific channels the connected account can already view.
VercelOct 21Vercel released durable-harness, an open-source package that makes coding-agent harnesses like Pi, OpenCode, Claude Code, and Codex durable. It writes each turn step to a durable log and uses queue backstops to resume sessions automatically after crashes or timeouts. This allows agents to run for hours or days without manual restarts or repeating completed work.
VercelOct 19Vercel added Inception AI's Mercury 3, a diffusion-based text model, to its AI Gateway at 50% off launch pricing. The model features a 1-million-token context window, tool calling, and structured outputs, with generation speeds exceeding 1,000 tokens per second. The model is accessible via the AI SDK or Vercel CLI using the inception/mercury-3 identifier.
VercelOct 19Vercel launched fx0, an open-source design agent that generates single-file HTML artifacts from natural language descriptions. Built on libfx, the agent features durable execution that resumes sessions from the last completed step if a function crashes. It integrates with Vercel AI Gateway to provide access to over 200 models for prototyping and design tasks.
VercelOct 19Vercel Shop is now generally available, providing an open source Next.js storefront template for Shopify. Built on the rebuilt Hydrogen SDK, the template exposes machine-readable product data and supports MCP tools for catalog search and cart management. It includes a shopping assistant powered by eve, allowing agents to browse and checkout directly at the merchant’s domain.
VercelOct 19Vercel added MiniMax's open-weight multimodal M3.1 Flash model to AI Gateway, free through October 22. The model features a 1-million-token context window and is optimized for coding and agentic tasks. To ensure free access, the provider must be pinned using gateway options, as the offer applies only when MiniMax serves the request.
VercelOct 19Vercel Labs added a snapshot --delta flag to the agent-browser CLI, which returns only changes since the last snapshot rather than the full page. This update reduces token usage by 73x during form-filling tasks. Because the delta resets between sessions, the tool supports saving snapshots to a baseline file for cross-session diffing.
VercelOct 18Vercel AI Gateway logs now group requests into sessions, surfacing each conversation's cost, cache hit rate, and the coding agent used. Session affinity automatically routes turns to the provider already holding the prompt in cache, improving efficiency. Session grouping is configurable via session IDs or headers; requests without explicit identifiers get a fingerprinted session automatically.
VercelOct 18Vercel added Elorian AI's Glimpse visual reasoning model to its AI Gateway, enabling image and video analysis, chart interpretation, and object tracking. The model is accessible via the elorian/glimpse identifier with one API key and no provider account. AI Gateway applies no markup or platform fees to inference requests, including those using bring-your-own-key authentication.
VercelOct 17Vercel updated its fx coding agent to support durable execution via a new libfx adapter API. The createFxAgent function now persists state across crashes, redeploys, and timeouts, allowing agents to resume mid-tool-call. This durability works on local machines or in the cloud using files, Postgres, or Vercel Workflows, providing production-grade reliability with one line of code.
VercelOct 16Vercel added Reflection AI's Beam to its AI Gateway, making the low-latency US open-weight model free for a limited time. Previously available in stealth as stealth/cat-jam, the 501B-parameter model supports a 1-million-token context window and tool calling. The model is accessible via the reflection/beam or reflection/beam-free identifiers using the AI SDK or OpenAI-compatible APIs.
VercelOct 16Vercel added Meta's Muse Spark Open, a US open-weight reasoning model, to its AI Gateway. The model features a 1-million-token context window and native text and image multimodality. Developers access the model via the AI SDK or OpenAI-compatible APIs, with built-in usage tracking, request tracing, and failover support without requiring a separate provider account.
VercelOct 16Vercel's October 2026 AI Gateway Index reports open-weight models now process over 70% of tokens and account for nearly 25% of spend. TypeSafe AI's Jev became the fastest-adopted model in gateway history, while Anthropic's share of total spend fell below 50% for the first time. Additionally, US-based providers captured $2.50 of every $3.50 spent on Chinese-origin models.
VercelOct 15Vercel Drives is now generally available, providing persistent storage for agents on Vercel Sandbox. The feature decouples file storage from execution environments, allowing agents to maintain state across runs and switch runtimes without data loss. Drives support forking and cross-sandbox mounting, with capacities ranging from 1 GiB on Hobby plans up to 16 TiB.
VercelOct 15Vercel launched early access to Token Pass, a $30 monthly subscription providing $80 in credit for GLM 5.3 and Kimi K3 models. The service routes requests through AI Gateway with Zero Data Retention enabled by default and inference in the US and EU. Access is limited to the first 500 Pro teams, with no automatic overage charges.
VercelOct 15Vercel Connect now supports MCP Events in public beta, enabling agents to subscribe to real-time webhook events from MCP servers. This update allows agents to act on triggers like new messages or document feedback without waiting for user prompts. The Connect SDK provides tools for managing subscriptions and verifying incoming event deliveries directly within application routes.
VercelOct 15Vercel published findings from its $1M sandbox security challenge, which drew 1,285 reports and resulted in $242,000 in bounty payouts. The program uncovered one Linux kernel zero day vulnerability, which Vercel mitigated in collaboration with industry partners. No reports showed evidence of customer data compromise, and independent architectural reviews confirmed the platform's virtualization boundary remains sound.
VercelOct 14Vercel added NVIDIA's Nemotron 3.5 Lightning to its AI Gateway, making the open-weight 30B reasoning model available for free. The model supports tool calling and structured outputs, targeting long-running agents, chatbots, and RAG systems. Developers access the model via the AI SDK or OpenAI-compatible APIs using the nvidia/nemotron-3.5-lightning-free identifier.
VercelOct 13Vercel added Microsoft AI's MAI-Image-2.6 and MAI-Image-2.6-Flash models to its AI Gateway. Both models support photorealistic imagery, text rendering, and targeted edits using up to five reference images. The AI SDK and AI CLI provide access via a single API key, with the 2.6 model optimized for final assets and the Flash variant for high-throughput production workloads.
VercelOct 13Vercel added Google's Gemini 4 Argon, the first Gemini 4 model, to its AI Gateway. The model supports a 1-million-token context window, multimodal inputs, and tool calling for complex coding and agentic tasks. It is available at 50% off through December 1st, with support for thinking levels and integration across the AI SDK and coding agents.
VercelOct 10Vercel reports that bot-originated traffic now accounts for 58.18% of its network, up from 32% in January 2024. Agentic deployments have surged to over 60% of the platform's total, compared to roughly 3% in January 2026. Additionally, agents now drive up to 83% of pageviews to Vercel’s documentation sites, a figure that rises with agent-optimized content.
VercelOct 10Vercel added Liquid AI's d1 decision model to AI Gateway, enabling classification, routing, and scoring by evaluating state against typed questions. The model returns structured, probabilistic answers without generating text tokens and now supports vision for image-based tasks. Developers access d1 via the AI SDK, OpenAI-compatible Decisions API, or TypeSafe-compatible API.
VercelOct 10Vercel added Microsoft AI's Decision-1 model to its AI Gateway, enabling text classification, request routing, and response scoring. The model returns structured answers with calibrated probabilities for choice, score, and boolean questions. It is accessible via the AI SDK, OpenAI-compatible, or TypeSafe-compatible APIs using a single AI Gateway key without requiring a separate Azure deployment.
VercelOct 9Vercel updated its CLI to include an agent-facing skill for domain management. Agents can now search for domains, check availability and pricing, and initiate purchases directly through the terminal. The system requires human approval to finalize the transaction, ensuring agents cannot spend money without explicit confirmation.
VercelOct 9Rillet runs eve agents on Vercel to convert customer requests into pull requests. This workflow allows the team to ship new features to production in as little as two hours. Since implementing these agents, the company has tripled its rate of merged pull requests.
v0Oct 9Vercel launched v0 for teams, converting personal accounts into shared workspaces for collaborative app development. The update adds real-time project viewing, shared chat access, and co-creation. It introduces granular privacy controls for chat visibility and a centralized dashboard to track team activity and project progress.
VercelOct 9Vercel added xAI's Grok Imagine Video 1.5 Lite to its AI Gateway, supporting video generation up to 1080p resolution. The model creates 1 to 15 second clips from text or image prompts with native audio. It supports seven aspect ratios and is accessible through the AI SDK and Vercel AI CLI.
VercelOct 8Vercel AI Gateway now supports OpenAI's Ultrafast service tier for GPT-6 Astra and GPT-6.1 Sol. This tier delivers faster output for interactive applications and coding agents, accessible by setting service_tier to ultrafast in the AI SDK or API. Ultrafast requests are billed at 6x the standard per-token rate and support US and global processing.
VercelOct 8Vercel added Black Forest Labs' FLUX 3 Image model to its AI Gateway, enabling image generation and editing through a single API key. The model supports up to 10 reference images for composition control and generates output at resolutions up to 4K. Developers access the model via the AI SDK or Vercel CLI without a separate provider account.
VercelOct 8Vercel added StepFun’s Step 5 Preview model to its AI Gateway, providing access via the stepfun/step-5-preview identifier. The 600B-parameter Mixture-of-Experts model features a 1-million-token context window and supports text and image input. Integration includes usage tracking, request tracing, and compatibility with coding agents without requiring a separate provider account.
VercelOct 8Vercel added Browserbase Search and Fetch tools to its AI Gateway, enabling any tool-calling-capable model to perform web discovery and page retrieval. These helpers, available in AI SDK 7.0.116 and later, provide consistent tool access across model providers using a single API key, maintaining functionality even when switching between different models.
VercelOct 8Vercel added Anthropic's Claude Haiku 5.5 to its AI Gateway. This model serves as the fastest Claude option at standard speed, optimized for subagents and summarization. It introduces adjustable effort levels, allowing control over reasoning depth and token usage, and supports Zero Data Retention for secure, high-volume tasks.
VercelOct 8Vercel added Glyph Cluster, a stealth reasoning model for coding and long-context analysis, to its AI Gateway. The model is free during its stealth period for Pro and Enterprise teams with purchased AI Gateway credits. It supports function calling and streaming, though it lacks zero data retention, and prompts may be used for model improvement.
VercelOct 7Vercel's AI Gateway now supports confidence-based decision fallbacks, allowing agents to escalate uncertain decisions to a secondary model. Users set a confidence threshold for Choice and Score questions, triggering a second decision when the primary model's certainty drops below the limit. This beta feature bills for both decision stages and operates alongside existing error-based failovers.
VercelOct 6Vercel added Google's Nano Banana 2.1 image model to its AI Gateway. The update delivers sharper text rendering, more consistent character generation across edits, and 4K support. The model is available through the AI SDK, the OpenAI-compatible Chat Completions API, and the Vercel AI CLI using the google/gemini-nano-banana-2.1 identifier.
VercelOct 6Vercel added Mistral AI's Mistral Large 4 to its AI Gateway, providing access to the 1T-parameter, natively multimodal model. The integration supports text and image reasoning through the AI SDK and OpenAI-compatible APIs without requiring a separate Mistral account. The gateway handles usage tracking, routing, and failover for the model using a single API key.
VercelOct 5Guillermo Rauch released gdp-ts, a library, linter, and AI skill for TypeScript that enforces compile-time authorization proofs. Sensitive functions now demand evidence of an authorization check, preventing missing checks or incorrect values. The linter blocks attempts to bypass these proofs, ensuring that coding agents cannot ship security bugs by skipping required checks in tight loops.
VercelOct 4Guillermo Rauch outlines a documentation strategy for his latest project, separating human-facing content from agent-facing internals. He hand-writes READMEs for human storytelling and connection, while using AI English for internal documentation intended for agent consumption. He notes that while AI-generated text is improving, it requires heavy human editing to be effective.
VercelOct 3Vercel integrated TypeSafe AI's Jev decision model into the AI SDK for Python, accessible via the experimental evaluate API. The release includes two experiments: a classifier that distinguishes Python code from English text in real-time, and a method for generating Python code through a sequence of structured decisions. Install via uv add ai.
VercelOct 2Vercel reports that Rogo AI engineers build internal applications on its platform, reaching over 73,000 deployments last month. Coding agents ship updates to production in five minutes, while agent swarms on the AI SDK manage production incidents without manual intervention. These six production agents automate workflows ranging from churn risk analysis to the sales team's deal desk.
VercelOct 1Vercel added Convai Innovations' Laya decision model to its AI Gateway, offering free access through October 31, 2026, when served by Boundless. The model evaluates context to return structured boolean, choice, or scoring answers. Integration is available via the AI SDK evaluation API, native HTTP, or TypeSafe-compatible APIs for routing agent tasks, triaging support, and enforcing guardrails.
VercelOct 1Vercel partnered with Microsoft AI to bring the MAI model family to its AI Gateway. The integration includes MAI-Voice-2.1 for long-form speech, MAI-Transcribe-2-Streaming for live transcription, and the newly released MAI-Voice-2.1-Flash for low-latency voice interactions. These models support zero data retention and are available for use through the AI SDK without platform markup.
v0Sep 29v0 integrates ChatGPT Plus and Pro subscriptions, enabling the use of existing ChatGPT tokens for AI generation within the platform. Connecting an account provides direct access to OpenAI models for building web applications. This integration utilizes the capacity of the linked ChatGPT plan for generation tasks within the v0 interface.
VercelSep 29Vercel added OpenAI's GPT-6.1 Sol to its AI Gateway, providing access via the openai/gpt-6.1-sol identifier. The model improves on the GPT-6 Sol added last week across coding, computer use, multi-step workflows, and complex document analysis. It features lower standard input and output pricing than GPT-6 Astra, with additional cost savings for requests that reuse long shared context.
VercelSep 28Vercel added Anthropic's Claude Sonnet 5.5 to its AI Gateway, reporting the model runs 30% faster and 30% cheaper than Sonnet 5 across benchmarks. It supports Zero Data Retention and integrates with the AI SDK, OpenAI-compatible APIs, and coding agents. The model excels at bug fixing, multi-step coding, and creating polished documents, slides, and spreadsheets.
VercelSep 28Vercel added Fireworks' Ember-1 model to AI Gateway, a research preview for coding and agentic workflows. The model delivers quality comparable to Kimi K3 while generating approximately 40% fewer tokens. It supports a 1-million-token context window, image input, tool use, and zero data retention. Ember-1 is available as a research preview for an initial two-week window.
VercelSep 26Vercel added the anonymous Pixel Canary model to its AI Gateway, offering free access while the model remains in stealth. Pixel Canary targets agentic coding and mobile app development, tying GPT-6 Astra (high) with a 90.3% success rate on Next.js evals. Prompts and responses sent through the model may be used for training and model improvement.
VercelSep 26Vercel reports that OpenAI's GPT-6 Sol now leads the DeepSecBench Pareto frontier, achieving the highest benchmark score while operating at an 80% lower run cost than the runner-up, GPT-6 Astra. The leaderboard ranks model configurations by balancing cybersecurity vulnerability detection accuracy against total execution cost.
VercelSep 26Klaviyo built an internal app platform on Vercel, allowing 512 employees to ship 356 projects in two weeks. The platform turns ideas into live, full-stack apps on Klaviyo’s own databases in three minutes. Every deployment remains private and SSO-gated by default, using Vercel’s Secure Compute to connect to internal infrastructure without crossing the public internet.
Frequently asked questions