Unified gateway to hundreds of LLMs via a single Vercel key. Zero token markup, BYOK support, automatic fallbacks, observability, and native AI SDK integration.

Vercel AI Gateway is a unified API gateway to hundreds of LLMs from multiple providers (OpenAI, Anthropic, Google, xAI, Meta, DeepSeek, Alibaba/Qwen, Mistral, Cohere, Nvidia NIM, Groq, Together, Fireworks, Perplexity, and others) served through a single Vercel key. The product was announced at Ship AI in May 2024, with general availability (GA) in 2025. Core value: a developer can switch between models and providers by changing model: 'anthropic/claude-sonnet-4.5' to model: 'openai/gpt-5-1' with no code changes and no need to create separate accounts with each provider.
Key features: (1) Zero token markup - Vercel adds no margin; token prices match provider prices exactly. (2) BYOK (Bring Your Own Key) - the user can use their own provider key through the gateway; tokens are billed directly by the provider, and Vercel provides routing/monitoring/observability for free. (3) Automatic fallback - if one provider fails, the gateway automatically reroutes to an alternate provider for the same model (e.g. Anthropic Claude via Anthropic โ Bedrock โ Vertex). (4) Provider options - explicit routing, provider preferences, a fallback list configurable per request. (5) Embeddings - support for vector embeddings (OpenAI, Cohere, Voyage). (6) Web search tool - augmenting model responses with real-time web search built into the gateway. (7) Observability - per-request metrics: model, provider, latency, cost, input/output tokens, aggregate spend dashboards.
API and integrations: AI Gateway supports three main API formats - (a) native Vercel AI SDK (open-source TypeScript, streamText, generateObject, useChat), (b) OpenAI Chat Completions API (drop-in replacement for clients using the OpenAI SDK - just change the base URL), (c) OpenAI Responses API (newer OpenAI API from 2025), (d) Anthropic Messages API (drop-in for Anthropic clients). Framework integrations: Next.js, React, Svelte/SvelteKit, Nuxt, Astro, Remix. Supported modalities: text-in/text-out, image-in (vision), image-out (generation), video-in/video-out (providers: Runway, PixVerse, Google Veo), embeddings.
Authentication and BYOK: the gateway supports two auth models: (1) OIDC access tokens generated by the Vercel account (recommended - short-lived, auto-rotated) and (2) classic API keys (longer static keys). BYOK lets the customer maintain a direct billing and contractual relationship with the model provider (OpenAI, Anthropic, etc.), while still benefiting from the Vercel layer for routing, observability, prompt caching, and disallow-prompt-training. Ideal enterprise scenario: OIDC + BYOK + opt-in prompt logging per project + custom per-team rate limits.
Market position and competition: AI Gateway competes directly with OpenRouter (both offer a unified LLM API, but OpenRouter runs a marketplace of 400+ models with a slight markup, while Vercel is zero-markup with BYOK), Portkey, Anyscale Aviary (more enterprise-focused), LiteLLM (open-source proxy). Vercel AI Gateway wins on native integration with Next.js/AI SDK (~3M weekly downloads), zero-markup pricing, and built-in observability in the Vercel console. Announced roadmap (Q3-Q4 2026): coding agents on top of AI Gateway, native RAG orchestration, wider multi-modal support.
Pricing models
Resource quotas
SLA & Support