Kyma API
Gives developers one key and one endpoint for a hundred open and frontier models.
The problem #
Developers who want open weight models end up holding a key and a bill per vendor. Every vendor fails in its own way, at its own hour, with its own error shape. What follows is not product work: it is retry code, key rotation and private uptime notes.
The approach #
One key, one OpenAI compatible endpoint, and a chain that retries the same model elsewhere. Reliability here is measured, not promised: every chat model gets probed every six hours. The result is published per model, next to the price, on a public page. The trade is price: one vendor bought direct will usually list cheaper than Kyma.
Where it is now #
Live since 2026-04-04, serving real traffic every day, with 102 models on one endpoint. An MCP door carries 18 tools, and the SDKs ship on npm and PyPI. Revenue is real but small, and top of funnel is still the binding constraint.
What happens if this works #
If buyers truly choose on reliability, gateways will publish per model uptime within three years. An unmeasured claim of 99.9 percent then stops being accepted, like a missing status page. Kyma is the test of that, because it already publishes that number on a public page.
Evidence #
- 3.19 billion tokens across 257,882 requests from 2,927 developers since 2026-04-05
- 18,637 requests (7.2 percent of all traffic) hit an upstream failure and still returned an answer; lifetime success 97.58 percent
- 102 models on one endpoint, each carrying a measured 30 day uptime rather than a claimed one
- every chat model probed every 6 hours under one uptime definition, published on the status page
- 69 paying customers since launch (cash figure not published here)
Notes #
Kyma started from a boring failure. A coding agent running against a single provider would stop mid task because that provider was rate limiting, or down, or had quietly renamed a model. The fix each time was the same shape: add another account, add another key, add another branch of retry code. After the third one it stopped being a fix and started being the product.
So the routing came first and the catalogue came second. A request for a model tries that model on another provider before it tries a different model, and it tells the caller what happened in the response headers. Over the lifetime of the service that has turned 18,637 failed upstream calls into answers the caller never had to know about.
The part that is harder to copy is the measurement. Every gateway in this category advertises 99.9 percent uptime and none of them say how they counted it. Kyma probes every chat model every six hours, counts a success only when the model that was asked for is the model that answered, and puts the resulting number on the public page next to the price. It costs a small amount per month to run and it is the one claim here that a buyer can check without trusting anyone.
What it has not solved is distribution. The service is reliable and the economics work per request, but not enough developers with a real workload have found it yet. That is the honest state as of September 2026.