Vercel

Vercel's AI SDK published @ai-sdk/tool-web-fetch 1.0.0, its first-party web fetch tool for agents. The release note is brief, so details of its behavior and limits are not stated in the release itself.

ByteDance's Seedance 2.5 lands on Vercel AI Gateway: single-call 30-second clips with synchronized audio, up to 50 mixed image/video/audio/style references, targeted edits that hold frame and pacing, and asynchronous generation that outlives serverless timeouts.

GLM 5.3 lands on Vercel's AI Gateway with better multi-step agent and software engineering results than GLM 5.2 at fewer output tokens, a 1M token context window, 128K max output, and a setup path into Claude Code, Codex and other coding agents.

Vercel's Sandbox SDK adds Linux users and groups: createUser gives each agent its own identity and private home directory that peers cannot read, write, or list, while createGroup opens a shared workspace for the cases where agents genuinely need to collaborate.

Tailscale's Aperture routes access to hundreds of AI models through Vercel AI Gateway and Sandbox, granting or revoking model access by tailnet network identity instead of per-tool API keys, cutting a custom routing build down to months.

Vercel Connect gains line-level visibility into the token lifecycle: every request, authorization, refresh, revocation and trigger delivery, linked by stable tokenId and authorizationId correlation IDs. Retention runs 12 hours to 30 days by plan, with drains for anything longer.

Vercel's new integration connects OpenAI's hosted Agents API loop to Vercel Sandbox, processing signed webhooks through Vercel Queues to give each agent session an isolated, persistent execution environment without long-lived VMs.

Vercel Connect reaches GA. Rather than provisioning a long-lived token per integration, code requests a short-lived scoped credential at runtime against its deployment's OIDC identity, across 100+ connectors, with RBAC and audit logs added for production use.

Vercel explains why cloud-hosted coding agent harnesses need durable execution, since a plain VM does not record completed model and tool calls or resume interrupted turns. It introduces eve and fx and argues for running non-sandbox work in functions.

An experimental vision-capable DeepSeek V4 Flash is now callable through the AI Gateway, accepting images alongside text in the same request, with tool use, reasoning and caching unchanged.

MiniMax H3 and H3 Max run at half price on Vercel AI Gateway from August 30 to September 13. H3 generates 2K video from text, an image, a first and last frame pair or reference media; H3 Max renders faster at 480p and 768p. Model IDs are unchanged, so existing requests get the rate.

Thinking Machines' Inkling Small is now callable on AI Gateway as thinkingmachines/inkling-small. It claims parity with the larger Inkling at about a quarter the size, reasons natively over audio and images, can crop and zoom to inspect fine detail, and supports Zero Data Retention routing.

Vercel's Is Agentic audits now group checks by site type (docs, business, app, commerce), so a commerce site sees payment/checkout standards like x402 while an API surfaces auth and SDK checks, set via a single meta tag.

Vercel AI Gateway logs now include a Sessions view grouping requests by session, with cost, models, providers, and cache hit rate per coding agent. Session affinity pins turns to one provider for cache reuse, using a sessionId or x-session-id header or a conversation fingerprint.

Vercel adds an official Cline adapter to the AI SDK harness layer, built with the Cline team, so Cline runs behind the same HarnessAgent interface as Claude Code, Codex, Deep Agents, Grok Build, OpenCode and Pi, with no bridge process installed inside the sandbox.

Vercel extended Spend Management, budget alerts, webhooks, and deployment pausing, to Enterprise teams on Flexible Commitment plans at no added cost, though AI Gateway and v0 usage remain unaffected by pausing.

Vercel Connect adds public beta MCP Events: apps and agents subscribe to webhook events from MCP servers, with automatic renewal, idempotent subscriptions and SDK verification of authenticated deliveries.

AI Gateway budgets now attach to a team or a project, not just an API key, and a request has to clear every budget it falls under — one scope over its limit rejects the call until it resets. Alerts fire at 50%, 75% and 100%, defaults cascade to unbudgeted projects, and the CLI manages all of it.

Vercel Connect adds a Managed API Key method for Stripe connectors, creating a separate permission-limited key per agent or task, stored by Vercel and rotatable without redeploying, with CLI support.

Vercel now charges Sandbox snapshot storage from each day's average usage rather than a single monthly average, with the daily charges visible on the team Usage page. The $0.08 per GB-month rate is unchanged and the switch applies automatically next billing period for Pro and Enterprise.

TypeSafe AI's Jev, a probabilistic decision model that returns typed choices and scores for use in agent tool-routing and guardrails, became the fastest-adopted model in Vercel AI Gateway history within 24 hours.

Chat SDK's Teams adapter picks up reactions and native ephemeral messages behind its existing cross-platform API, so an agent can acknowledge work with an emoji and ask permission where only one user sees the prompt. Author emails now arrive via Microsoft Graph, and token factories replace static se

Chat SDK's new chat/workflow subpath turns human approval into one call: requestApproval posts an Approve/Deny card and suspends a Workflow SDK run for seconds or days, surviving deploys. Approvers are scoped, clicks are signature-verified, and the card edits in place as an audit trail.

Grok 4.6 lands on Vercel's AI Gateway: 500K-token context, text and image input, four reasoning levels defaulting to high, selectable as xai/grok-4.6 in the AI SDK or inside Claude Code, Codex, OpenCode and Pi after one setup command.