New API vs LLM Gateway

Why

The licences and the token economics are identical, which is what makes the rest of the comparison legible. Both are AGPL-3.0, both charge zero markup on tokens, both charge $0 per seat, and both ship a console for users, keys and spend rather than being bare proxies. Deployment is where they part. New API is self-host only — Docker Compose for production, a single container, 1Panel and BaoTa one-click installs, and a cluster mode that scales behind an external Nginx or HAProxy — with no vendor-hosted tier on any page and no vendor in the request path. LLM Gateway offers both: the whole platform self-hosts free forever from one Docker image, or you use the hosted service and pay a flat 5% when you buy credits, plus 1.5% on international cards, with BYOK free on the free plan and no seats, minimums or subscription. If you were never going to hand your traffic to a vendor, that 5% is not a cost you would ever pay, and New API's deployment story is the more complete one.

Community scale is the headline asymmetry and it buys less than it looks like. New API has 47,090 stars, 11,229 forks and 1,318 open issues, was last pushed 2026-09-01, and is unambiguously active — but the shipping line is still v1.0.0-rc.30, published 2026-08-31, after nearly three years, its docs changelog trailed at rc.25 when checked, and the project's own documentation says it is primarily intended for personal learning and research, that stability is not guaranteed, and that official technical support may not be provided under the open-source licence. LLM Gateway has 601 stars, is operated by Polar Lights LLC, a Delaware company that publishes no funding or ownership information, and released v1.14.0 on 2026-08-24 on a weekly minor cadence. It also carries SOC 2 Type II, independently examined, an enterprise DPA and a status page. New API publishes no SOC 2, no ISO 27001, no HIPAA BAA and no GDPR DPA at all, and its acceptable-use policy pushes identity management, log retention and filing duties onto the deployer. On paper, the small project is the one you can put in front of procurement.

Infrastructure and integration follow the same line, and this is where LLM Gateway earns its place. It publishes an official Helm chart as an OCI artifact on GitHub Container Registry deploying the gateway, API, UI and worker against PostgreSQL and Redis, Terraform modules on the Enterprise plan that provision EKS, RDS, ElastiCache, networking and secrets, a hosted MCP server at api.llmgateway.io/mcp exposing chat, image generation and model listing to Claude Code, Codex and Cursor, and a first-party Vercel AI SDK provider. New API documents none of that: no Kubernetes or Helm, no Terraform, no MCP, no OpenTelemetry, and routing configuration lives in the console database rather than in Git. What it has instead is API surface. One New API instance serves three dialects — an eleven-endpoint OpenAI surface covering chat, completions, embeddings, image generation and edits, audio transcription and speech, rerank, responses, a realtime WebSocket and models, plus native Anthropic Messages and native Gemini v1beta — while LLM Gateway documents OpenAI Chat Completions and the Anthropic Messages endpoint, with embeddings, audio, batch and the Responses API not documented on the pages checked. On counts, LLM Gateway states 200+ models across 40+ providers against New API's 100+ models and 30+ providers, both sides publishing floors rather than counted lists.

Then there is what each does with governance, and both answers are partial. New API's strength is access control rather than content policy: tokens are first-class virtual keys with expiry, remaining quota, an unlimited-quota flag, model restrictions and an IP allowlist, auto-disabling when the quota runs out, with group multipliers, three-tier ratios at 1 USD to 500,000 quota points, pre-consumption reconciled after the call, and EPay, Stripe, Creem and Waffo top-ups, redemption codes and subscription plans around it. What it does not publish is any PII, injection or custom-policy engine — only a blocked-word facility whose matching mode, default state and returned error are all undocumented. LLM Gateway advertises guardrails, PII redaction and content moderation enforced at the gateway, but the governance tier is where the open core bites: audit logs, region pinning and per-project routing customisation are Enterprise, code in the repository's ee/ directory needs a separate commercial licence, and enterprise features require a signed licence in production with a seven-day grace period once it expires. New API's AGPL exemption is likewise a commercial licence arranged by email at an unpublished price. Retention differs in kind rather than degree: New API has no vendor-side copy at all and a Log Retention Days setting you choose, while LLM Gateway defaults to metadata-only with full payload retention billed at $0.01 per million tokens.