Best LLM Gateways of 2026

Portkey is the production-gateway pick. Founded in 2023 in San Francisco. The wedge is full feature breadth: caching, fallbacks, prompt management, guardrails, RBAC, SSO, and observability across 40+ providers. Portkey open-sourced their full gateway under Apache 2.0 on March 24, 2026.

Open Source Apache 2.0 is free forever with Universal API, retries, routing, guardrails, fallbacks, and load balancing for self-host. Developer Free hosted covers 10K logs a month with key management and basic observability. Production at $49 a month covers 100K logs with unlimited prompt templates, semantic caching, LLM guardrails, and RBAC; overages bill at $9 per 100K up to 3M (the realistic SMB paid entry). Enterprise covers 10M-plus logs, VPC hosting, SSO, SOC 2 Type II, GDPR, and HIPAA.

The trade-off versus OpenRouter: Portkey ships the gateway features that matter at production scale (cache reduces token spend; fallbacks survive provider outages; guardrails enforce content policies) but indie-dev brand recognition is narrower. OpenRouter wins the head-term reader by name; Portkey wins the production-team reader by feature breadth plus the open-source self-host option.

OpenRouter is the api-aggregator and the indie-dev brand reference for LLM gateways, founded in 2023 in San Francisco. The wedge is uniquely-true: 300+ models from 60+ providers behind a single OpenAI-compatible endpoint with pay-as-you-go credit-balance pricing and pass-through provider rates (zero markup on inference tokens).

Free gives every account a $1 starter credit with full access to all models for evaluation. Pay-as-you-go bills consumed tokens at the underlying provider price with no inference markup; a small Stripe processing fee applies on credit purchases. Pro at $20 a month is a credit minimum that unlocks higher rate limits and Slack support; the realistic indie-dev paid entry. Enterprise at $1,000 a month covers dedicated routing, private deploys, an SLA, and audit logs.

The trade-off is feature breadth, not pricing. OpenRouter ships fallbacks and basic rate limiting but not caching, guardrails, or prompt management; for high-volume production workloads the same workflow on Portkey Production at $49 a month adds caching that often pays for itself in token savings. OpenRouter is the right call when model variety matters more than gateway feature breadth and when you do not want to manage your own provider keys.

Vercel AI Gateway is the frontend-stack-native pick. Launched in 2024 by Vercel (founded 2015 in San Francisco). The wedge is uniquely-true: bundled with Vercel platform plans and tightly integrated with the AI SDK v5/v6 ecosystem; no token markup; hundreds of models behind a unified endpoint with BYOK support.

Hobby is free with the Vercel Hobby plan with hundreds of models, no markup, BYOK, and built-in observability. Pro at $20 per user a month is bundled with Vercel Pro and adds load balancing, fallbacks, budgets, and spend monitoring; the realistic developer paid entry. Enterprise covers private clusters, SAML SSO, audit logs, and SLA.

The trade-off is ecosystem lock-in. Vercel AI Gateway is excellent if your stack is already on Vercel and you use the AI SDK; the integration is the cleanest in the category and tokens cost the same as direct provider pricing. Outside the Vercel ecosystem, the gateway's value drops because the AI SDK integration is the load-bearing differentiator. Teams not on Vercel should pick OpenRouter or Portkey instead. Composite math wins not by feature breadth but by zero markup and the bundled price.

LiteLLM is the open-source-MIT-proxy pick. Founded in 2023 by BerriAI Inc (Y Combinator W23) in San Francisco. The wedge is uniquely-true: Apache 2 / MIT-licensed Python proxy that translates 100+ provider APIs to the OpenAI Chat Completions schema. Run it as a self-hosted service in front of your model providers and your application code targets the OpenAI SDK only.

Open Source is Apache 2 / MIT-licensed and free forever with self-hosted Python, 100+ provider integrations, and community support. Cloud Free is a hosted tier with limited request volume. Cloud Pro at $50 per user a month covers cost tracking, budgets, team management, and virtual API keys; the realistic team paid entry. Enterprise covers self-hosted enterprise, SSO, SOC 2, and HIPAA available.

The trade-off versus closed gateways: LiteLLM is the right tool when self-host control matters (regulated industries, on-prem deploys, code transparency) but the OSS Python proxy requires Python operational maturity. Many teams run LiteLLM behind their own application proxy and get the same routing benefits as Portkey or Helicone without the per-request cost. Cost-conscious teams at scale often migrate from OpenRouter to self-hosted LiteLLM.

Helicone is the observability-first pick. Founded in 2023 (Y Combinator W23) in San Francisco. The wedge is uniquely-true: one-line drop-in proxy that logs requests, tracks cost per session/user/property, and ships an HQL query language for analytics. Set the base URL once and every OpenAI SDK call gets logged with full request/response, latency, and token cost.

Hobby is free with 10K requests a month, 1 GB storage, and a single seat. Pro at $79 a month covers unlimited seats, alerts, reports, and HQL; the realistic SMB paid entry. Team at $799 a month covers 5 organizations, SOC-2, HIPAA compliance, and dedicated Slack support. Enterprise covers custom MSA, SAML SSO, on-prem deployment, and bulk discounts.

The trade-off versus full gateways: Helicone is observability-led with caching and fallbacks added later; the primary use case is logging and cost analytics, not routing. The Pro tier repriced from $20 to $79 a month in 2025 (a 295 percent increase). Compared to Langfuse Core $29 a month, Helicone Pro is more expensive but ships gateway proxying alongside the observability. Choose Helicone when you want observability AND gateway in one product.

Cloudflare AI Gateway is the edge-native pick. Launched in 2023 by Cloudflare (founded 2009 in San Francisco). The wedge is uniquely-true: gateway bundled with Cloudflare Workers across the 300+ data center edge network. Calls run from the closest Cloudflare PoP to the user with cached responses served at the edge.

Free covers 10K requests a day with caching, analytics, rate limits, and OpenAI/Anthropic/Workers AI provider support. Workers Paid at $5 a month covers higher request limits and edge-native gateway access; the realistic developer paid entry. Workers Standard at $200 a month bundles 50M Workers requests with unlimited AI Gateway calls. Enterprise covers dedicated regions, private connect, SLA, and audit logs.

The page score uses Workers Standard $200 because the tier name contains the substring that catches the standard match; this is a 3,900 percent overshoot from the realistic Workers Paid $5 entry, the largest single overshoot in this category and one of the largest in the system. The cons block acknowledges the gap. Cloudflare AI Gateway is the right call when you already run Cloudflare Workers and edge-native gateway latency matters; outside the Workers ecosystem, OpenRouter or Portkey ship more gateway features.

Langfuse is the open-source-observability pick. Founded in 2023 (Y Combinator W23) in Berlin Germany. The wedge is uniquely-true: MIT-licensed OSS LLM observability with self-host option; tracing, evals, prompt management, and analytics on a project that you can run on your own infrastructure or use as Cloud.

Hobby is free with 50K observations a month, all platform features (with limits), 30-day data retention, and 2 users. Core at $29 a month covers 100K observations with unlimited users and 90-day retention; the realistic team paid entry. Pro at $199 a month adds 3-year retention, SOC 2, ISO 27001, HIPAA, and 20K req/min throughput. Enterprise at $2,499 a month covers audit logs, SCIM API, dedicated support, and SLAs.

The trade-off versus Helicone: Langfuse is observability-only (no gateway proxying) and ships as MIT OSS so you can self-host for unlimited usage. Helicone is observability-plus-gateway in one product. Choose Langfuse when self-host matters and gateway routing belongs to a separate layer (LiteLLM as proxy, Langfuse as observability). The Pro tier overshoots Core ($199 vs $29) by 586 percent because the heuristic matches the Pro tier name; the realistic team budget is Core $29 a month.