Final Router

Changelog

What shipped, and when

Every change a customer would notice, newest first. A gateway that stops changing is a gateway falling behind the models it routes to - so this page is also the honest answer to whether anyone is still building here.

  1. New

    Provision keys through the API

    Create, cap, rename and revoke gateway keys programmatically - for platforms issuing a key per customer, per tenant or per environment. It authenticates with a new kind of credential: a management key administers keys and cannot reach a provider or spend a cent, so leaking one is not the same as leaking the other. Create one from the API keys screen.

    Read more

  2. New

    OpenAI's Responses API

    POST /v1/responses accepts OpenAI's newer request shape and answers in it, streaming included - so code written against the Responses API can point here without being rewritten, and reach every model in the catalogue rather than one vendor's. It is stateless: previous_response_id and store are refused by name rather than accepted and ignored.

    Read more

  3. New

    Anthropic's Messages API

    POST /v1/messages speaks Anthropic's Messages format, authenticated with x-api-key exactly as their SDKs send it. A Claude-native codebase changes one base URL and keeps working - except the model field can now name any of the seventeen models here. Routing, fallback, guardrails and billing are the same code as the chat endpoint; only the dialect differs.

    Read more

  4. New

    See what routing saved you

    Analytics now totals a figure that was previously recorded per request and shown nowhere: for every request the router chose, the difference between the dearest model it could have used and the one that answered, priced at the same token counts. Requests where you pinned a model count zero - that was your decision, not the router's.

    Read more

  5. New

    Cost attribution, documented

    Label requests by app, end customer, session and tag, then read spend split along each of them. Every response already carried its cost in cents; the page now explains how to make that decompose into causes, and the request log exports exactly the filtered view as CSV for reconciliation.

    Read more

  6. Improved

    Embeddings and the prompt library get their pages

    Both have been running for a while and appeared on no page. Embeddings run through the same key, caps and log as chat, at list price with no markup. Stored prompts are callable as prompt/name with checked variables, so fixing a system prompt is a save rather than a deploy. The API reference also now documents the embeddings, cost-estimate and cache-flush endpoints.

    Read more

  7. Improved

    Stronger signup and payment protections

    New defences against the stolen-card pattern that targets AI gateways: additional checks at signup, limits that ease as an account establishes itself, and automatic detection of coordinated payment abuse. Legitimate accounts should notice nothing beyond a brief ceiling in the first day, which lifts by itself.

  8. New

    Governance: roles, approvals and an audit trail

    A workspace member can be promoted to admin - managing gateway keys and guardrails while billing, provider credentials and the roster stay with the owner. Members can request a key and a second person must approve it, with the requester creating the key so the secret is only ever seen by whoever holds it. Every role change and decision lands in a trail nobody can edit.

    Read more

  9. Improved

    Choose how long prompts are kept

    Prompt storage is still off unless you switch it on. When it is on, you can now pick 1, 7, 14 or 30 days instead of always 30 - and shortening the window deletes anything already older than it, immediately, rather than waiting for the next sweep.

    Read more

  10. New

    Bring your own keys, structured output and fallback

    Three capabilities got the pages that explain them: your own provider keys with the fee published rather than quoted, json_schema working on every model through a forced-tool translation, and cross-provider fallback that names every model it tried in the response.

    Read more

  11. New

    A public MCP server

    Connect your coding assistant and it can look up live models, capabilities, prices and provider health while it writes your integration - instead of hardcoding a guess that drifts. Four tools, public catalogue data only, no key required to read.

    Read more

Older changes predate this page. For the shape of the whole product, start at the features or the API reference.