--- title: "Lazu vs New API" description: "Lazu is based on New API. What was kept, what was rewritten (routing, billing, usage records, protocol conversion), and which one to run." source: https://lazu.ai/docs/compare/new-api updated: 2026-11-03 --- # Lazu vs New API Lazu is based on [New API](https://github.com/QuantumNous/new-api), which is based on [one-api](https://github.com/songquanpeng/one-api). The lineage is kept in the repository's NOTICE and in the console, as the upstream's licence terms ask. This page is about what changed, not about which is better: New API is a well-run project with a large community, and anyone who already operates it has no reason to migrate unless the differences below matter to them. Facts were checked on 2 November 2026. ## The shared base Both are Go gateways with a web console. Both accept OpenAI Chat Completions and Responses, Anthropic Messages and Gemini `generateContent` requests, keep upstream channels with priorities and weights, retry on failure, manage users, groups and API keys with quotas, and bill in a prepaid balance. Both run from Docker Compose with PostgreSQL and Redis. If you know one console, the other is familiar. ## What Lazu rewrote **Billing is a contract, not a multiplier.** New API prices with ratios and expression-based pricing and stops a token or user at `403 insufficient_user_quota`. Lazu replaced the charge path with a written [billing contract](https://github.com/lazu-ai/lazu/blob/main/docs/billing-contract.md): funds are reserved before the call, settled from the provider's usage counters, and released when a request fails without billable output. Each usage dimension records whether the provider measured it, Lazu estimated it, or it is unavailable; an unavailable number is never filled in from the reservation, and an account is never silently overdrawn. Prices are listed per model and per lane in a catalog the key can query. **A usage record the customer can read.** New API keeps usage and audit logs for the operator. Lazu exposes each completed request to the key that made it: normalized usage, measurement source, billing line items and timing, fetched by request ID ([request details](https://lazu.ai/docs/endpoints/usage-requests)). Requests from a coding agent group into sessions, and [a script](https://github.com/lazu-ai/lazu/blob/main/ops/scripts/summarize-agent-session.py) reconciles an exported session offline. **Protocol conversion with a fidelity rule.** New API converts between its inbound formats and provider formats. Lazu rebuilt this on an immutable request representation: each attempt plans its route from the original request, a conversion is performed only along edges verified lossless, and a request that needs a feature the upstream cannot carry is refused with a `400` naming the feature, before the call ([design note](https://github.com/lazu-ai/lazu/blob/main/docs/protocol-conversion-architecture.md)). **Routing you can see.** Channel selection by priority and weight, retries and taking a failing channel out of rotation are shared. Lazu keeps client errors from counting as channel faults, and publishes a route-health page so users see the state of each route without asking the operator. **Teams and projects.** New API has users and groups. Lazu adds projects with members, roles, invitations and per-member budgets, so one organization can run several teams on one balance. **Extension points.** Features a deployment compiles in attach through documented hooks ([docs/EXTENSIONS.md](https://github.com/lazu-ai/lazu/blob/main/docs/EXTENSIONS.md)); the core builds and runs without them. New API extends through JavaScript task plugins for media tasks and otherwise through source changes. **A product, not only a gateway.** Lazu ships its public pages, a model catalog with lanes and prices, and a docs site in English, Chinese, Traditional Chinese and Japanese, generated from one source with Markdown and `llms.txt` for agents that read docs. ## What New API has that Lazu does not - A larger community, more channel types and faster coverage of new providers. - Subscriptions and redemption codes, and payment integrations beyond Stripe. - Installers for 1Panel and Baota, and multi-node cluster docs. ## Side by side | | New API | Lazu | | ------------------- | -------------------------------------------- | ----------------------------------------------------------------------------------------------- | | Base | one-api | New API | | Licence | AGPL-3.0 with attribution terms | AGPL-3.0-or-later; keeps New API's and one-api's attribution | | Billing | Ratios and expressions; quota stop at `403` | Reservation → settlement → release, per-dimension measurement source, prices per model and lane | | Usage records | Operator logs | Per-request record readable by the key; agent sessions; reconciliation script | | Protocol conversion | Converts inbound formats to provider formats | Lossless edges only; `400` naming the feature otherwise | | Routing | Priority, weight, retry, affinity | Same, plus a public route-health page | | Organizations | Users and groups | Users, teams and projects with member budgets | | Extending | JavaScript task plugins; source changes | Go and web extension points | | Docs | docs.newapi.pro | lazu.ai/docs in four languages, with `llms.txt` | ## Which one to run Run **New API** if you already run it, if you need its subscription and redemption features, or if you want the widest channel coverage as soon as a provider ships. Run **Lazu** if what you sell or hand out is an account: budgets that are reserved before the call, records the customer can read, a catalog with prices, and a public route page. See also [Lazu vs LiteLLM](https://lazu.ai/docs/compare/litellm) and [the overview of other gateways](https://lazu.ai/docs/compare/overview).