The best LLM gateways in 2026, scored on what they actually do

The comparison table

Scored on what the product does today, from vendor documentation and repository metadata read in August 2026. "Good" means it is a reason to choose this product; "basic" means present and shallow; "no" means absent.

The three most people should pick from

LiteLLM: the open-source default

The most-adopted open-source gateway by a wide margin, 56,700 GitHub stars in August 2026, and the only one where "does it support provider X" is almost always yes. The core is MIT-licensed; the enterprise/ directory in the same repository is under a separate commercial licence, which is why GitHub reports the repo's licence as unresolved. That split is honest but it does surprise people who assumed the whole thing was MIT.

What you get free: virtual keys, teams, spend tracking, budgets and rate limits, fallbacks, request logging, Prometheus metrics. What Enterprise adds, at a price quoted per gateway capacity rather than per token and never published: SSO and SCIM, OIDC and JWT auth, audit logs, secret-manager integration and key rotation, multi-region control plane, and 24/7 SLAs.

OpenRouter: the fastest path to many models

One account, one key, hundreds of models, and no infrastructure. If a provider errors, OpenRouter falls back to the next provider serving the same model automatically. Routing shortcuts are pleasant: append :nitro to a model slug to sort providers by throughput, :floor to sort by price, and openrouter/auto hands model selection to their classifier.

The cost is a fee and a middleman. Credit purchases carry 5.5% (minimum $0.80) by card and 5% by cryptocurrency, checked against the OpenRouter FAQ in August 2026. Bring-your-own-key usage is fee-free up to a monthly allowance of list-price inference (documented as $25,000/month on pay-as-you-go) and 5% above it. In exchange, every prompt you send crosses a third party that is not the model vendor.

Portkey: the governance pick

Portkey's gateway is MIT-licensed and self-hostable with no paywalled core routing, and the hosted platform adds guardrails, prompt management, and access control. Pricing read from the Portkey pricing page in August 2026: Developer is free with 10,000 recorded logs a month and three prompt templates, Production is $49/mo for 100,000 logs with overage at $9 per additional 100,000 up to three million, Enterprise is custom and quoted at 10 million plus logs with SSO, VPC hosting, and SOC 2 Type 2.

The reason to choose it over LiteLLM is that guardrails and role-based access are first-class rather than bolted on, and the reason not to is that the free tier's 10,000 logs a month is a demo allowance, so the real comparison is $49/mo against your own operations time.

The six that win narrower cases

• Helicone is an observability product with gateway features, not the other way round. Hobby is free with 10,000 requests a month, 1 GB of storage and 7-day retention; Pro is $79/mo with unlimited seats and one month of retention; Team is $799/mo with five organisations, SOC 2 and HIPAA. Ingestion is rate-limited by tier, from 10 logs per minute on Hobby to 15,000 on Team. Choose it when the thing you cannot see is why a prompt regressed, not when you need routing.
• Bifrost (Apache 2.0, 7,400 stars) is the performance-first entrant, written in Go and marketed on gateway overhead. The claims are the vendor's own and worth benchmarking rather than believing, but the licence is clean and the code is readable. Smaller community than LiteLLM, so expect to read source.
• Kong AI Gateway is a set of LLM plugins on Kong Gateway (Apache 2.0, 44,000 stars). Right answer if Kong already fronts your APIs and your security team already knows it. Konnect, the managed control plane, is the paid path and starts around $105/month per gateway with a million requests included, which secondary sources agree on and Kong does not publish in a form you can quote confidently.
• Vercel AI Gateway charges no markup and no platform fee on tokens, including with your own keys, which is the most aggressive pricing position in the category. The add-ons are metered separately: custom reporting at $0.075 per 1,000 tag writes and $5 per 1,000 report queries, team-wide provider allowlist and team-wide zero data retention at $0.10 per 1,000 requests each. Compelling on Vercel, pointless off it.
• Cloudflare AI Gateway gives you dashboard analytics, caching, and rate limiting free on any Cloudflare account, with persistent logs capped at 100,000 total on Workers Free and 10 million per gateway on Workers Paid. Unified Billing, where Cloudflare pays the providers for you, carries a 5% fee on purchased credits. Thin on routing, excellent as a free front door.
• Continuum is the odd one out and belongs here only because people arrive at this page from a coding-agent problem rather than an application one. Its gateway is fused to a workbench that runs Claude Code, Codex, Cursor, Grok and OpenCode, with an auto model router, per-repo cost analytics, and organisation-level model policy and weekly spend caps. BYOK is free forever; hosted inference is $25/mo and up. It is not a general-purpose proxy you would put in front of a production API, and if that is what you need, one of the eight above is the answer.

How to choose in ten minutes

Questions people ask

Keep reading

Continue through LLM gateways and routing with LLM routing: how a router decides which model answers.