What shipped on the gateway, newest first. Dates are the days the changes went live.
Platform
Legal and Security pages list every place content is stored
The Privacy Policy, Data Protection page and Security page now name the four time-limited places a prompt or response can be written to our database: Capture (1, 7 or 30 days), the response cache (up to 24 hours, with an embedding for the semantic variant), the Batch API (30 days) and Idempotency-Key retries (24 hours). They also state that request logs hold the classified provider error line and any user, HTTP-Referer and X-Title values you send, and that the semantic cache sends the conversation to the embedding provider.
Billing
Batch API bills at standard per-token pricing
Batches no longer carry a discount: each item is billed at the same per-token price as a synchronous request to the same model, and the batch object reports discount_pct 0. The 24-hour window, the 500-request limit and 30-day result retention are unchanged.
Reliability
Uptime probe and outage hardening
An external probe now watches the gateway from outside our own network and raises an alert the moment inference stops answering. Gateway failures return clean JSON instead of proxy error pages, the service restarts itself until it is healthy, and the marketing site falls back to a baked catalog snapshot instead of ever reading as empty when the API blinks.
Models
Independent quality rankings
Model pages and the rankings board now carry quality data we did not produce ourselves: LMArena human-preference ratings with their confidence intervals, and Epoch AI benchmark scores. Every number names its source, license and snapshot date, so you always know which figures we measured and which we did not.
ModelsReliability
More routes to the same models
Cohere and Perplexity adapters joined the gateway, along with the HuggingFace router, which brings fourteen upstream hosts behind a single key. When several providers serve the same model, the gateway routes each request to the cheapest healthy one and fails over automatically, and the model page shows every route with its price.
PlatformBilling
Batch API: 500 requests, half price
POST /v1/batches accepts up to 500 inline chat requests, runs them within a 24-hour window, and bills them at half the synchronous price. Results are kept for 30 days.
ModelsPlatform
Embeddings endpoint
POST /v1/embeddings is live, priced per million tokens like everything else, with embedding models listed in their own section of the catalog.
Platform
Bring your own provider keys
You can now attach your own OpenAI, Anthropic, Google and other provider keys. Requests on your keys are billed by the provider directly and carry the same platform fee as a credit top-up. Stored keys are encrypted with AES-256-GCM, verified against the provider before saving, and only ever shown back as their last four characters.
Platform
Web search, Responses API and response caching
Three gateway features in one wave: an OpenRouter-compatible web search plugin that grounds any model with live results, the OpenAI Responses API at POST /v1/responses, and an opt-in response cache that replays identical requests at zero cost, scoped to your own account.
ModelsBilling
Free tier and the auto router
42 chat models plus the first speech and image models became callable free with the :free suffix, backed by a daily sponsorship pool with a public meter. rapidrouter/auto picks a sensible model for you when you do not want to choose.
BillingPlatform
Checkout and invoices
Checkout went live with an invoice on every payment. Per-key spend caps, account budgets, model allowlists and outbound webhooks landed in the same wave.
Models
Eleven more providers, one public catalog
Eleven additional upstream providers came online behind the gateway, and the full priced catalog went public: every live model with its price, context window and capabilities, browsable without an account.
Platform
Real accounts and live provider status
The demo login is gone. Sign-up, sessions and API keys run against the real backend, only models that are actually callable are listed anywhere, and the providers page reads its status live from the router itself.
Platform
Rapid Router goes live
rapidrouter.ai launched: one OpenAI-compatible endpoint in front of the major AI models, priced per million tokens, pay as you go with no subscription.