Your LLMs all speak OpenAI now, and it tracks the receipts

What it does

new-api is a Go service that sits between your applications and upstream AI providers. It exposes a single interface—mostly OpenAI-compatible—and handles protocol translation for Claude, Google Gemini, DeepSeek, and others. Beyond proxying, it adds organization-level auth, token quotas, weighted channel routing, and per-request cost accounting with cache-hit billing.

The interesting bit

The project treats API format fragmentation as the main enemy. It cross-converts between OpenAI, Claude Messages, and Gemini formats bidirectionally, and abstracts reasoning-effort controls across o3-mini, GPT-5, Claude 3.7 Sonnet, and Gemini 2.5 by appending suffixes like -high or -thinking to model names. That is a lot of string-packing for a gateway.

Key highlights

• Supports OpenAI Responses, Realtime API, Claude Messages, Gemini, and rerank endpoints (Cohere, Jina) under one roof.
• Built-in billing logic for usage-based and cache-hit accounting across OpenAI, Azure, DeepSeek, Claude, and Qwen.
• Authentication options include OIDC, Discord, Telegram, and LinuxDO login.
• Database-compatible with the original One API project for easier migration.
• Can proxy non-chat workloads like Midjourney image generation and Suno audio generation.

Caveats

• Gemini-to-OpenAI conversion is text-only; function calling is not yet supported.
• OpenAI Responses format conversion is still marked as in development.
• The README repeatedly emphasizes legal compliance and authorized use, suggesting operators need to navigate regulatory overhead carefully.

Verdict

Worth a look if you run a multi-tenant or multi-provider AI setup and need a single billing and routing layer. Skip it if you only call one API and don’t need quota management.