Vercel AI Updates: October 16, 2026

1. Vercel Drives Brings Persistent Agent Storage to Vercel Sandbox

Vercel. Vercel made Drives generally available for Vercel Sandbox, giving agents persistent storage that is separate from the execution environment so an agent can move to a new sandbox with different tools, runtimes, or compute and keep working on the same files. A drive is created once and mounted at a path (for example /workspace) through the CLI or the Sandbox SDK’s Drive.getOrCreate and mounts option, and only files under that path persist. Multiple sandboxes can mount the same drive, read-only copies can be shared across agents, and forks create independent writable copies that do not merge back. Drives are available on all Sandbox plans with a 1 TiB default size (1 GiB on Hobby) and up to 16 TiB. Source

2. Token Pass Bundles $80 of GLM 5.3 and Kimi K3 Credit for $30 a Month

Vercel. Vercel introduced Token Pass, a subscription in early access for Pro teams that costs $30 per month and includes $80 of monthly credit for GLM 5.3 and Kimi K3, with more models planned. It runs through AI Gateway and works with coding agents such as fx, Command Code, and Pi or with custom agents built on the AI SDK. Zero Data Retention is on by default with inference in the US and EU, unused credit does not roll over, and requests stop when credit runs out rather than incurring overage charges. Source

3. AI Gateway Production Index: Open-Weight Models Take 71.5% of Tokens

Vercel. Vercel’s October AI Gateway Production Index reports that open-weight models handled more than 70% of gateway tokens in September and nearly one in four dollars of spend, up from about one in 20 dollars in June. DeepSeek-v4.1-Flash processed nearly half of September’s tokens but only 2.6% of spend, and more than two-thirds of spend on Chinese-origin models went through US-based providers. Anthropic’s share of gateway spend fell from 63.7% to 43.9%, while Opus 5.5 took 11.03% of spend in its first two weeks. Nearly one in two teams changed their most-used model in September, and gateway spend was about 13 times December’s level as average price per token fell 47%. Source

4. Vercel Connect Adds MCP Events in Public Beta

Vercel. Vercel Connect, the service that links apps and agents to external MCP servers, now lets them subscribe to webhook events from compatible MCP servers in public beta. Connect creates and automatically renews subscriptions on the MCP server, verifies incoming webhooks, and forwards events through authenticated delivery, with an SDK to confirm a delivery came from Connect. An MCP Events adapter exposes event discovery and subscription as AI SDK agent tools, and in the eve framework enabling events on an MCP connection gives agents tools to create, inspect, or stop subscriptions. Source

Vercel. Vercel Connect added Enterprise-Managed Authorization (EMA) in public beta, allowing apps and agents to access supported services on behalf of signed-in users without a separate consent prompt for each service. Administrators grant or revoke access through an identity provider such as Okta, and Connect handles token exchange, caching, and renewal within approved scopes and the user’s existing permissions. Apps using Vercel Passport can, for example, request a token for Slack’s MCP server using the signed-in user’s identity, and Connect also supports Cross-App Access (XAA) for MCP authorization. Source

6. Ship 2026: eve Gains Persistent Memory and AI Gateway Adds Budgets

Vercel. Vercel’s Ship 2026 recap lists updates to eve, its open-source agent framework, including persistent memory across sessions, extensions for installing tools and skills, and agent creation and deployment from the Vercel dashboard. AI Gateway now supports budgets for API keys, projects, individuals, and teams, along with US-only inference and Zero Data Retention. Vercel also said its internal support agent automates 91% of support tickets and that 45% of questions to its eve-based data agent d0 now come from other agents. Source

7. Meta’s Muse Spark Open Available on AI Gateway

Vercel. AI Gateway added Meta’s open-weight Muse Spark Open as meta/muse-spark-open, a model aimed at coding and agentic workflows such as multi-file refactoring, debugging, and large-codebase understanding. It accepts text and image input with a 1M-token context window and is offered with Zero Data Retention and no training. The model is available through the AI SDK, OpenAI-compatible Chat Completions and Responses APIs, the Anthropic-compatible Messages API, and coding agents connected to the gateway. Source

8. Liquid AI’s d1-ultra Decision Model Comes to AI Gateway

Vercel. AI Gateway added Liquid AI’s liquid/d1-ultra, a multimodal decision model for classification, routing, and scoring over text and images that returns Boolean, Choice, or Score answers with probabilities instead of generated text. Several questions can be asked about the same input in one request, and the context window is 32,768 tokens. Billing is on input tokens only (including images), and the model is accessible through the AI SDK’s experimental_decide, an OpenAI-compatible Decisions API, and a TypeSafe-compatible API. Source

9. TypeSafe AI’s Jev 1.14 Adds Image Input and Long-Context Mode

Vercel. AI Gateway now serves TypeSafe AI’s Jev 1.14 as typesafe-ai/jev-1.14.0, a decision model that takes state plus a set of questions and returns typed Choice, Score, and Boolean answers with probabilities in one parallel request. Version 1.14 accepts an image as a file part in state and adds an experimental long-context mode enabled with providerOptions.typesafe.longContext: true for long documents, logs, or conversation histories. Zero Data Retention and no-training can be enabled per request. Source

10. Stealth Model Cat Jam Free on AI Gateway

Vercel. AI Gateway added stealth/cat-jam, an unnamed-provider reasoning model for coding and agentic tasks designed for low-latency inference, free during an unspecified stealth period. It supports function calling, JSON-schema structured outputs, returned reasoning text, and reasoning effort levels from low to max, with text-only input. Zero Data Retention is not available, and prompts may be retained for safety reasons. Source