LiteLLM vs Vercel AI Gateway: LLM Gateways Comparison

LiteLLM is an open-source AI Gateway developed by BerriAI with 18,000+ GitHub stars, enabling unified access to 100+ LLM APIs through OpenAI-compatible format. Founded as a Y Combinator company with USD 1.6 million in seed funding, LiteLLM is trusted by companies like Rocket Money, Samsara, Lemonade, and Adobe. The platform provides retry and fallback logic, cost tracking, guardrails, and load balancing with MIT licensing for the core proxy. While the open-source version is free, running LiteLLM requires infrastructure costs of USD 200-500 monthly plus DevOps labor, monitoring tools, and incident response. The Enterprise version at USD 30,000 annually adds SSO, RBAC, and team-level budget enforcement. Users praise LiteLLM's unified API interface and security through open-source auditability, but note production complexity with latency overhead (20-40ms) and operational burden for self-hosting.

Vercel AI Gateway is a production-ready LLM gateway that provides unified access to hundreds of AI models with built-in reliability, monitoring, and cost management. Available to every Vercel team account, the gateway offers a free USD 5 monthly credit plus pay-as-you-go pricing with zero markup on model tokens. When bringing your own API keys, Vercel charges no platform fees, offering tokens at provider list price. Vercel Agent is priced at USD 0.30 per action plus underlying token costs. The platform focuses on developer experience, enabling frontend developers to add LLM capabilities with minimal setup without managing provider-specific SDKs or credentials. Key features include unified API across providers, budget controls, usage monitoring, load balancing, and automatic failover. While praised for ease of use, transparent pricing, and reliability with automatic failover, Vercel AI Gateway faces criticism for infrastructure limitations including 504 Gateway Timeout errors for long-running agents, execution time constraints (15 seconds default, 300 seconds maximum on Pro), insufficient semantic caching relying only on HTTP headers, and vendor lock-in with custom middleware not easily portable to other platforms.