LLM Gateway

LLM Gateway (llmgateway.io by Polar Lights LLC) is a unified API for routing requests across 40+ LLM providers, including OpenAI, Anthropic, Google, Meta, Mistral, and DeepSeek-so teams use one OpenAI-compatible endpoint and one key for 200+ models, with usage analytics, cost tracking, caching, and automatic failover. It is also known as an AI API gateway, LLM proxy, and unified LLM API. The product is open source under AGPLv3 and self-hostable; the managed cloud is operated by Polar Lights LLC (16192 Coastal Highway, Lewes, DE 19958, United States). Public interviews and the co-founders’ writing identify Luca Steeb and Ismail Ghallou (smakosh) as builders; the About page itself names the LLC and a distributed GitHub community rather than individual founders. As of September 2026 the GitHub repository theopenco/llmgateway shows on the order of 1.5K stars (API retrieve ~1,594 on 2 September 2026).

Mechanically, applications point an existing OpenAI SDK (or raw HTTP) at https://api.llmgateway.io/v1 (or a self-hosted base URL), authenticate with a gateway API key, and keep model strings interchangeable. The gateway routes to the chosen provider, meters tokens and cost per model/provider/project/key, caches repeatable responses, and fails over when an upstream degrades. Pricing for the gateway product is not a seat subscription: pay-as-you-go prepaid credits bill at provider list token rates plus a flat 5% platform fee at credit purchase ($10 minimum / $5,000 maximum per top-up; plus 1.5% for international cards per pricing.md last updated 15 August 2026). Bring-Your-Own-Keys (BYOK) is free with a 0% platform fee; self-hosting the AGPLv3 stack is free. Free models are available on the free tier without buying credits. Enterprise is custom (volume discounts, SSO, audit logs, guardrails, 99.9% uptime SLA, 30-day production pilot). Related Polar Lights surfaces-DevPass coding plans and Lounge chat memberships-are separate SKUs and must not be confused with gateway credit math.

What LLM Gateway delivers as of September 2026

The job-to-be-done is multi-provider LLM access without multi-SDK glue. Official docs describe the same loop: one integration, cost visibility across providers, reliability via failover and caching, and no lock-in because the stack is OpenAI-compatible and self-hostable. The public catalog at llmgateway.io/models lists supported LLMs with pricing, context windows, and capabilities; providers are enumerated at /providers. Marketing and pricing copy consistently claim 200+ models and 40+ providers-catalog counts should be verified against live listings before final procurement. New models are typically listed within about 48 hours of provider release (home/About).

Developer surfaces beyond chat completions include embeddings, vision, image and video generation paths (emphasized on the Portkey comparison page), Anthropic Messages compatibility, OCR/document helpers, realtime/speech/transcription endpoints in the OpenAPI surface, an MCP server with OAuth well-known metadata, an AI SDK provider integration, and agent/skills templates. Machine-readable docs ship as OpenAPI at api.llmgateway.io/openapi.json and concatenated markdown at llmgateway.io/llms-full.txt; the homepage and pricing page honor Accept: text/markdown for agent scraping. Rate-limit 429 responses include Retry-After and RateLimit-* headers. The API is path-versioned under /v1.

Official About page volume metrics: 100B+ tokens routed across 20M+ requests. Official About page statistics provide the verified baseline. Enterprise marketing also cites SOC 2 Type II compliance (report request via security.llmgateway.io), GDPR-aligned DPA/subprocessors, TLS in transit and AES-256 at rest including provider keys, and status at status.llmgateway.io.

Where LLM Gateway sits versus OpenRouter, LiteLLM, and Portkey

Against OpenRouter, the clearest difference is self-hosting and fee shape. LLM Gateway’s own compare page positions itself as open/flexible from $0 with free forever self-host, flat 5% on credit purchases, and free BYOK; it describes OpenRouter as a hosted proxy with 5.5% platform fee on card credits and BYOK that becomes fee-bearing above published thresholds. LLM Gateway publishes a migrate-from-OpenRouter guide: same OpenAI-compatible API and model-string pattern-swap base URL and key. Choose OpenRouter when you want the largest hosted model marketplace and do not need prompts inside your own network; choose LLM Gateway when AGPLv3 self-host, free BYOK, or a lower advertised credit fee matter.

Against LiteLLM, both are developer-centric proxies you can run yourself. LiteLLM is widely used as a self-hosted translation/proxy layer with optional enterprise support (LiteLLM lists enterprise tiers starting around $250/month on its own materials; re-check current LiteLLM quotes on LiteLLM’s site). LLM Gateway packages gateway + dashboard + worker as a full AGPLv3 platform with a managed cloud that charges 5% on credits (0% BYOK). Pick LiteLLM if you already standardize on its proxy config and only need self-host translation; pick LLM Gateway if you want one vendor path from free self-host through managed credits into enterprise SSO/guardrails without assembling the control plane yourself.

Against Portkey (no Vantaige slug confirmed in the local slug list at write time-link when imported), LLM Gateway’s compare page frames Portkey as a gateway + LLMOps suite (free tier, then paid plans; fees blog cites Production ~$49/month plus log overages) with broader claimed model counts, while LLM Gateway stresses full-platform AGPLv3 self-host, automatic provider scoring on live metrics, transparent 5%/0% BYOK pricing, and image/video through the same API. Portkey often wins teams buying deep prompt management and OpenTelemetry-centric LLMOps; LLM Gateway wins teams buying open self-host plus unified multi-provider routing with pass-through token rates.

Helicone and Kong AI Gateway sit in the same broad “AI gateway” category but did not have dedicated LLM Gateway compare URLs as of 2 September 2026-avoid inventing feature matrices without dedicated compare URLs. Helicone’s job is primarily observability and logging in front of LLM calls; Kong’s AI Gateway job is API-platform control for AI traffic inside Kong’s broader gateway ecosystem. LLM Gateway’s job is unified multi-provider model access with OSS self-host and credit/BYOK commercial paths. Vercel AI Gateway appears on LLM Gateway’s compare hub for teams already inside Vercel’s AI SDK ecosystem.

"One OpenAI-compatible API for 200+ models across 40+ providers. Per-token prices match each provider's published rates; LLM Gateway charges a flat 5% platform fee when you buy credits. Bringing your own provider keys is free, and self-hosting is free under AGPLv3." - LLM Gateway Pricing (pricing.md), last updated 2026-08-15

Cost, quotas, and real usage friction

There is no mandatory monthly gateway seat for the core API. Friction shows up as (1) the 5% credit uplift and card/international fees when you are not on BYOK, (2) the $10 minimum top-up, (3) self-host infra and on-call cost if you leave the managed cloud, and (4) enterprise commercial negotiation when you need SSO, SIEM-ready audit logs, guardrails, white-label, or a 99.9% SLA. Refunds on credit top-ups (and on DevPass/Lounge payments) are self-serve within 14 days if less than 20% of the purchase was used; DevPass Reset Passes differ (7 days, unused). Free-tier free models avoid credits entirely for experimentation but are not a substitute for frontier paid models.

Do not price the gateway using DevPass ($29 / $79 / $179) or Lounge ($9 / $19 / $49)-those are adjacent products for coding-tool flat usage and consumer chat. Token Cost Calculator and Copilot cost calculator pages on the marketing site help compare pass-through token spend to other billing shapes; always re-check llmgateway.io/models for live per-model rates before locking a budget.

Where LLM Gateway still breaks: and who should skip it

Catalog breadth is marketed as 200+ models / 40+ providers; OpenRouter and some LLMOps vendors advertise larger hosted catalogs-teams that need every long-tail hosted model may still prefer a marketplace-first proxy. AGPLv3 self-host is a feature for control and a constraint for companies that refuse copyleft without a commercial license for the ee/ enterprise features. Managed-cloud buyers who never use BYOK will always pay the 5% credit fee versus negotiating direct provider discounts alone. Observability depth versus specialist tools like Helicone, and API-management depth versus Kong, remain different product centers of gravity.

"LLM Gateway adds full self-hosting under an AGPLv3 license, deeper real-time cost and latency analytics for every request, free Bring Your Own Keys, and flexible enterprise add-ons. OpenRouter is a hosted proxy only and cannot be run on your own infrastructure." - LLM Gateway vs OpenRouter FAQ, llmgateway.io/compare/open-router, retrieved September 2026

Who should use LLM Gateway: and who should skip it

Use LLM Gateway if you need one OpenAI-compatible integration across many providers; if free BYOK or AGPLv3 self-host is a procurement requirement; if you want managed credits with an explicit flat platform fee and no token markup; or if you are migrating off OpenRouter and can swap base URL/key. Enterprise fits orgs that need SAML/OIDC SSO, per-member budgets, guardrails, audit logs, white-label chat, Terraform-based deploy, and a 30-day production pilot before a long contract.

Skip LLM Gateway if you only call a single provider SDK and do not need failover or unified billing-talk to that provider directly. Skip the managed credit path if a specialist observability gateway (Helicone-class) or an API-platform AI gateway (Kong-class) is the actual job and multi-provider model routing is incidental. Skip expecting a ChatGPT-style consumer chat product from the gateway API alone (Lounge is the separate chat membership). Skip inventing “unlimited free frontier models”-free models exist, frontier tokens still cost money at provider rates.