Vercel AI Updates: May 20, 2026

1. Gemini 3.5 Flash lands on Vercel AI Gateway with parallel agentic execution

Vercel. Google’s latest Flash model is live on AI Gateway under the google/gemini-3.5-flash identifier, advertising improved coding proficiency, parallel agentic execution loops, and tighter multi-turn coherence over prior Flash generations. The model defaults to medium thinking level, and provider options expose thinking adjustments and thought inclusion, but temperature, topP, topK, and thinking_budget are not supported on this surface yet. For AI SDK users this means a one-line model swap to compare Flash 3.5 against Sonnet 4 or GPT-5 for agentic workloads without leaving Gateway’s failover and observability tooling. Source

2. Vercel Sandbox now hosts Anthropic’s Claude managed agents inside Firecracker microVMs

Vercel. Teams can now execute Anthropic’s managed agents inside Vercel Sandbox, with each agent session pinned to its own Firecracker microVM and deny-by-default egress governed by a domain allowlist. Anthropic owns the model, harness, tools, and session state, while Vercel runs the execution environment and brokers credentials so secrets never enter the sandbox. The pitch is low-latency connectivity to private networks and internal services without exposing API keys or building a bespoke isolation layer, useful for teams already standardised on Claude that want to keep tool execution on their own infrastructure. Source

3. AI SDK OpenAI provider adds first-class speech and transcription model support

Vercel. @ai-sdk/openai@4.0.0-canary.62 introduces speech and transcription model support on the provider, extending the AI SDK’s modality surface beyond chat, embedding, and image generation into audio in and audio out. The canary release lands the wiring for the SDK’s experimental speech/transcription APIs against OpenAI’s gpt-4o-transcribe and TTS endpoints, so applications building voice agents or call-summary pipelines can now use the standard generateSpeech and transcribe helpers against OpenAI without dropping to raw fetch. The change ships on the 4.x canary track, separately from the stable 3.x line. Source

4. Flat Rate CDN enters limited beta for Pro teams with predictable monthly pricing

Vercel. A new Flat Rate CDN offering is in limited beta for Pro teams, replacing variable per-request and per-GB pricing with a predictable monthly cost that covers edge requests and data transfer without metered overages. The pitch is bill protection during traffic spikes - viral launches, AI-generated demand surges, or scraping waves no longer translate into runaway invoices, which has been a recurring complaint from teams running heavily-cached marketing sites and AI app frontends on Vercel. Pricing and quota specifics are gated behind the beta enrollment form. Source

5. Consolidated commit status collapses monorepo project checks into one GitHub signal

Vercel. Vercel now consolidates multiple project deployment statuses into a single GitHub commit check for monorepos, so a repo with ten linked Vercel projects no longer pollutes pull request status panels with ten green ticks per commit. Branch protection rules can target the one consolidated check instead of enumerating every project, which simplifies required-status configuration when projects come and go. The change is opt-in at the team level and lands without altering the underlying per-project deployment events that automation depends on. Source