Vercel

Vercel's AI SDK for Python adds an experimental evaluate() API for Jev, a model that answers multiple-choice questions with confidence. The post gives install steps, AI Gateway setup and a full example, and cautions that accuracy needs testing.

Every AI Gateway request now emits an OpenTelemetry span, exportable to any OTLP/HTTP endpoint with native Braintrust, Sentry, Dash0, Kubiks and Statsig targets. Spans carry model routing, retries, tokens, cost and TTFT — never prompt or completion text. Sampling is configurable; $0.05 per 1k traces

A benchmark from Vercel that measures how accurately different LLMs (GPT-5.6, Claude Opus 5, Kimi K3, Grok 4.5, etc.) find real security vulnerabilities in application code, using a secret, human-judged golden set of 231 findings across 50 entry-point files to prevent memorization. Each model is scored on a recall-weighted F2 metric alongside cost and runtime, revealing that even the best model catches only about 31% of known vulnerabilities while cheaper open-weight models increasingly rival frontier models on cost-efficiency.

Vercel's AI SDK harness layer now lets Claude Code, Cursor, Codex, and GitHub Copilot authenticate via native subscription logins instead of API keys, with OAuth tokens resolved and refreshed on the host across a shared harness interface.

Vercel's Security Dashboard is GA on all plans, ranking misconfigurations by risk and linking each to the setting that fixes it. The companion `vercel security check` CLI exposes findings an agent can remediate — enabling fork protection, marking env vars sensitive, swapping static credentials for O

Generated code that queries a warehouse should never hold the user's OAuth token, because prompt injection can steer it into exfiltrating whatever it can read. The fix is a request proxy outside the sandbox that resolves the real credential at request time.

Vercel's AI Gateway now exposes Exa search as a tool that works with any gateway model and needs no separate Exa key — called from the AI SDK via gateway.tools.exaSearch() — and it becomes the default web_search backend for eve agents.

Vercel put functions on Fluid, the same compute layer already running its builds and sandboxes, sized per workload and reconfigured in flight. It reports 15 million builds a day, 25 million sandboxes a week and a trillion requests a month, with agent sandbox startup as the motivating case.

A managed connector that registers the identity application and federated credentials for you, so an app or agent can request short-lived tokens for a large productivity suite without stored client secrets.

AI Gateway now exposes OpenAI's Ultrafast service tier for GPT-6 Astra. It is requested via a provider option, billed at 6x standard rates, limited to US and global processing, and works best with Responses API over WebSocket for tool-heavy loops.

Vercel CLI can now deploy small static artifacts (up to 10 HTML/Markdown files under 5MB) in under a second by skipping the build step, useful for publishing a prototype or agent-generated page instantly.

Vercel's monthly report on AI Gateway traffic finds open-weight models jumped from 11% to 29% of token volume between April and June 2026 while consuming under 4% of spend, with DeepSeek alone reaching 22.6% of tokens, just behind Google. Anthropic still dominates spend at 61% despite only 32% of token volume, and the report breaks down per-modality leaders (OpenAI's GPT Image, xAI's Grok Imagine) plus a brief note on Claude Fable 5's release and export-control suspension.

Vercel's CLI gained a traces search command that filters spans by status, service, deployment or request ID, supports a KQL subset for duration queries, and emits JSON per span for scripts and agents.

Vercel's AI SDK team built a "software factory" that triages issues and authors PRs autonomously while humans keep control of every merge. Four weeks in it authors 25-35% of merged PRs and closes 70-80% of issues, with the design argument centred on reviewer efficiency rather than generation speed.

Vercel Agent is now reachable in Slack for Pro and Enterprise teams in public beta. Mention it in a thread and it answers with deployment, build, log and metric context, investigates incidents, reviews pull requests, and executes approved changes under a temporary grant.

Apps and agents built in v0 can now reach more than 100 third-party services through team-owned connectors, with each connection minting a short-lived token rather than storing a long-lived secret.

Vercel's AI Gateway now routes zai/glm-5.3-flash: text and image input, a 1M-token context window, function calling, structured output and streaming. The coding-agents setup command wires it into Claude Code, Codex, Cursor, OpenCode and others.

Vercel expands Sandbox, its code-execution environment for agents, from 4 to all 20 compute regions, with configurable failover regions and SDK/CLI support for picking region at creation time.

Vercel's AI Gateway index for July: average price per token fell 13.6% even as spend grew 37%, DeepSeek overtook Google as the second-largest lab by volume, and Anthropic took 65% of spend on 30% of tokens. Kimi K3 arrived heavy, at about twelve times K2.5's tokens per request.

Vercel's AI SDK harness layer adds an official Cursor adapter, connecting it through the Agent Client Protocol so Cursor, Claude Code, Codex, Cline, OpenCode and others all run behind the same HarnessAgent interface.

Mem0 is now a native Vercel Marketplace integration, giving agents persistent cross-session memory with auto-provisioned API keys billed through your Vercel invoice.

Vercel's CTO argues the defender's advantage is temporary. Kimi K3 is a near-frontier open-weight model with no meaningful cybersecurity safeguards, topping open weights on DeepSec Bench, while most frontier models will already do defensive work — and teams are underusing them.

Vercel deploys are now up to 33% faster for apps using Incremental Static Regeneration, achieved by shipping routing metadata for prerendered pages in a single upload instead of splitting it into a separate one when it grows large. The optimization applies automatically to all deploys, including those using Partial Prerendering, with no configuration required.

Chat SDK gains an Instagram adapter so bots can handle DMs, media, quick replies, reactions and story replies through Meta's messaging API, with buffered streaming and a 24-hour reply window.