At 2,000 requests a second, LiteLLM dropped 8% of them.

Measured on LiteLLM v1.104.2 at stock settings, against the same upstream on the same host with pinned cores. Full method, every scenario and what the run does not show: the benchmark in full

What your renewal is buying

LiteLLM is migrating its gateway to Rust and markets the result hard. The migration is real. These three facts are theirs, published by them:

• Rust became the default for OCR in v1.102.0-rc.1, enabled per process and disabled with LITELLM_RUST=0.
• The published Rust figures measure the forwarding path, not a feature-complete gateway.
• The migration is staged through December 2026, which is to say it is not finished.

When the Rust path covers budgets, spend logs and audit, we will measure it again and publish whatever we find, including if it beats us.

You are paying for features that are Apache-2.0 here.

Multi-tenancy. Budgets below the key. Guardrails scoped per team. Secret manager integration. Their enterprise documentation lists all four as things a licence key turns on. In Wäg they are in the Apache-2.0 build, and they always have been.

Every capability we compared, including the ones we gate too, with the source for every row: the full licence boundary

Their agent gateway registers agents. It does not govern them.

LiteLLM shipped an A2A agent gateway and the basics are there: register an agent, control which teams and keys reach it, attribute the cost. What their documentation does not describe is a tool call held for a named human who has to say yes, a run that cannot call an outbound tool once it has read private data, a tool definition pinned so a silent change is caught rather than executed, or a credential that lives for minutes instead of forever. Their emergency stop fires a webhook at the agent's own runtime and leaves it registered and enabled on their side — it asks the thing you have lost control of to please stop. Ours refuses it at the door, on every replica and region, and cuts a stream that is already running.

What agent governance looks like · the agent layer, row by row

Keep Claude Code exactly where it is.

Your developers keep the provider-native path — Claude Code on your Anthropic plan, Codex, Gemini CLI. They set a base URL and one header with their Wäg key, and Wäg authenticates the caller, meters the spend against their team's budget and writes the request log. You stop having a blind spot in the one place your most expensive users work.

• Native endpoints and native SDKs, so each tool behaves the way its own documentation says it does.
• With the Wäg key header, identity, budgets and request logs apply against the same hierarchy as every other call. Without it, calls are still forwarded and metered, but no Wäg budget applies.
• With Enterprise, your guardrail chains run on that traffic too — the same scanners you already configured, on the one path that would otherwise go unscanned.

Stated plainly, because it is in our documentation anyway: on the open-source build, pass-through traffic flows unscanned and says so in a response header. Enterprise is what closes that gap.

Three ways in

Pick the one that describes you. None of them costs anything.

What moving actually involves

You keep your SDK, your model aliases and your application code. One base URL changes. Your LiteLLM config.yaml imports, and we tell you exactly what did not come across rather than silently dropping it.

The migration, step by step · Where Wäg and LiteLLM actually differ