6 posts tagged with "LLM gateway"

MLflow is the open-source ML/GenAI platform originally built at Databricks, and by far the largest, oldest project we've compared AcruxCore against — 27,000+ GitHub stars, 444 contributors, shipping since 2018. Its GenAI side is newer: a Prompt Registry, a tracing store, LLM-as-judge evaluation, and — the real surprise of this comparison — an AI Gateway that sits in the request path, the same design AcruxCore uses. We built the same prompt — vip-support-triage, a support agent that changes tone for VIP customers and lists their open tickets — on both platforms, then ran the identical sequence on each: create the prompt, version it, send a live call through a gateway, inspect the trace, build a dataset, and call it from an SDK script.

Every latency number we'd published before this measured one thing: the completion call. But a real request to any of these platforms is usually two round trips — fetch or resolve a stored prompt, then complete it — and most of them were benchmarked against OpenRouter, not the provider whose name is on the model.

So this one measures the full cycle, on real OpenAI billing, across all six platforms plus both of AcruxCore's calling modes — eight paths in one interleaved run. We also ran it four independent times, because a single 100-round run against a live third-party API turned out to not be enough to trust a single number, and that's worth showing rather than hiding.

Helicone is a proxy-first LLM observability tool — you point your API base URL at it and it sits in the request path, which makes it architecturally closer to AcruxCore's own gateway than most competitors in this series. So we built the same prompt — vip-support-triage, a support agent that changes tone for VIP customers and lists their open tickets — on self-hosted Helicone, then tried to run the identical sequence we've run on every platform in this series: create the prompt, send a live call, inspect the trace, build a dataset, and call it from a script. On this self-hosted build, three of those steps hit real, reproducible errors on Helicone's side before we ever got a trace. We're publishing those errors, not routing around them — that's the most useful thing we found.

A gateway virtual key can be given a cache window: set cacheTtlSeconds and an exact-match repeat of a call — same model, same messages, same temperature/max_tokens/top_p/stop — is served from Postgres instead of the provider. We ran the same fixed set of prompts through it repeatedly to see what a hit is actually worth, in milliseconds and in dollars.

Your app sends one POST /gateway/chat/completions request with an OpenAI-shaped body; changing providers is a one-field edit to model. What that field hides is more interesting: underneath, Anthropic, OpenAI, and Gemini each expect a genuinely different request — a different auth header, different sampling-param nesting, and a different key for the same tool definition. We took one fixed request — a system message, a user message, and one tool definition — and traced it through all three adapters that ship today to show exactly what each one rewrites and why the gateway needs to do it at all.

Putting a gateway in front of your model providers buys you a lot: one endpoint for every provider, cost accounting, caching, virtual keys, and budgets. But it raises an obvious worry — am I paying for that with latency? Every request now takes an extra hop, and for a user-facing app, milliseconds matter.

So we measured five ways to reach the same model — through our gateway, around it with your own key, and raw — to see exactly where the milliseconds go.