What shipped

The site now leads with the position we actually hold: Gatewayz is the inference layer for the agentic economy — built for software that calls a model thousands of times with nobody watching. Each claim on the page links to the page that backs it.

Why it matters. You can tell in a few seconds whether this is the right layer for what you are building.

Crawlers and answer engines now receive the real text of the homepage in the first response, instead of an empty shell that only fills in once scripts run.

The questions people were asking us by email now have a page each, with a straight answer on it. Quickstart gets you from an empty terminal to a working call; Security & data says what we keep, for how long, and what we never store.

How it works walks a single request from your agent to the provider and back. Integrations lists the tools we have tested end to end and is honest about the ones we have not. Solutions covers coding agents, agent platforms and autonomous organizations.

A point-by-point page on calling a provider's API directly versus calling it through Gatewayz — including the cases where going direct is the better choice, such as a single vendor and a hard latency budget.

Why it matters. A comparison that never concludes against itself is not worth reading.

A Start here path takes a newcomer from what inference is to a first working request, across ten foundation guides. Eight in-depth articles go deeper on routing, caching, spend ceilings, streaming failures and honest benchmarking, and a glossary defines the terms plainly.

The site publishes an llms.txt so AI crawlers can read a structured summary of what Gatewayz is, which errors are terminal and where the authoritative model list lives. Every static page is now prerendered, so what a crawler reads is what a visitor reads.

We removed claims across the site that we could not source or that were out of date: model counts, availability and latency figures, and savings percentages. Automated checks now fail the build if any of them come back.

Why it matters. Nothing on this site should be a number you have to take on faith.

Pricing is stated as the provider's list price plus a routing fee, and the per-model prices on the page are read from the live catalog rather than typed into the page.

The Benchmarks page sets out how we measure — gateway against direct, back to back, reported at p50 and p95 — and the rules that stop us publishing a latency claim without a direct baseline or a cost claim from a run with no cache reads. No production run is published yet, and the page says so.

Why it matters. You can judge the method now, and check the numbers against it later.

The public status endpoint now reports on models in the live catalog only, using measurements from the last 24 hours. A model with no recent measurement is reported as unmonitored rather than counted as down, and availability is reported as unknown rather than estimated.

Why it matters. A status page that guesses is worse than one that admits a gap.

The Claude Code instructions now give the base URL without a trailing /v1 — with it, the SDK builds a doubled path — and name the auth-token variable the CLI actually reads. Links that had gone stale across the docs now resolve.

GET /health/quick reports the exact build serving requests, and the status page shows it.

Why it matters. After a deploy you can confirm from the outside which build answered you, rather than assuming.

If a provider fails partway through a streamed response, the stream now ends with an error event instead of simply stopping.

Why it matters. A truncated answer and a complete one used to look identical to the client reading them.

An undated vendor alias now resolves to the exact dated snapshot behind it, and an id that is not in the catalog returns 400 model_not_found. We resolve aliases. We never substitute a different model for the one you asked for.

An unknown model id returns 400 model_not_found rather than a 503 that SDKs quietly retry. A spent per-key request cap returns 402 request_cap_exhausted, and an empty balance returns 402 insufficient_credits, in place of a retryable 429.

Why it matters. An unattended agent retries on 429 and 5xx. A condition that retrying cannot fix now says so in the status code, before it costs you a budget.

A request that presents an API key is now either authenticated with that key or refused. It is never carried on as an anonymous request.