What's new

Every release, and what it changed

Newest first. Anything that changes how the gateway behaves is listed here, including the things that were taken away.

19 September 2026 · 0.1.0-alpha.6

A vendor key carries what it may be used for, and every refusal is recorded

  • Each of your vendor keys now has its own model list. One key can serve the cheap models and another the expensive ones, which is what an organisation with a development key and a production key actually wants. Narrowing a list tells you first which virtual keys would lose which models.
  • A vendor key can be tagged development, staging or production. A production virtual key can then only be routed through a production vendor key, and the editor refuses the mismatch when you save it rather than at three in the morning.
  • A vendor key can carry a monthly spend cap of your own. Past it, requests through that key are refused — or fall back to the organisation default, if that is what the virtual key asks for.
  • Every refused or failed request is recorded: which key, from which address, with which client, when and why, and never the secret. A new Refusals page says what needs a look — a revoked key still being called, someone trying wrong secrets, a key used from an address it has never worked from — and which of your limits are actually being hit, how often and by how much.
  • A key is suspended automatically after five different new addresses in an hour, rather than after a count of attempts. One address retrying is a broken client; five new ones is a key in circulation. Wrong secrets never suspend a key, however many are tried.
  • “Pin to the first address used” became a checkbox under the address list: add the next new address automatically, once, within ten minutes. It adds exactly one address and unticks itself.
  • The tokens-per-minute limit on a key is gone. A cheap model returns a great many tokens for very little money, so the number never bounded what anyone set it to bound; requests per minute, budgets and the per-request cap do.
  • The key wizard was rebuilt: the route is chosen before the models, and each step shows what you have chosen as you go.

19 September 2026 · 0.1.0-alpha.5

Embeddings, speech and transcription, and a re-read catalogue

  • Embeddings through OpenAI, speech through ElevenLabs and OpenAI, transcription through AssemblyAI, ElevenLabs and OpenAI — all on the OpenAI-compatible paths.
  • Your organisation can name ElevenLabs voices, so an application sends voice: “narrator” and the voice behind the name can change without a deploy.
  • Every model in the catalogue now declares what it does, so a model sent to the wrong endpoint is answered with a 400 naming the right one instead of “no such model”.
  • Prices were re-read from the vendors' own pages. Four were wrong, one of them charging five times too little; corrections are new price rows, so past requests keep the price they were charged at.

17 September 2026 · 0.1.0-alpha.4

Image generation

  • POST /v1/images/generations through OpenAI's image models, off by default on every key and switched on per key with ceilings on how many pictures, how large and at what quality.
  • Every generated picture is stored with the request that made it, because a picture the gateway produced has to be accountable for.

Every update is included

Buy the product once and every later release is yours.