Speko (YC S26) Launches OpenRouter for Voice AI
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Speko launched in Y Combinator's S26 batch with a routing platform that picks optimal STT, LLM, and TTS combinations based on user constraints — accuracy, latency, cost, or balanced — and returns the provider and model names in response headers.
- Founder Bek spent four years building enterprise voice agents across Asia in 10+ languages before founding Speko, repeatedly running the same ritual of hiring native-speaking raters and benchmarking each new speech model against the existing production stack.
- Speko publishes its benchmark boards publicly, including cases where its selections perform worse than alternatives, and does not train or sell models itself to keep rankings impartial; failover happens only during connection setup, falling back to runners-up if a provider refuses the attempt.
- The company open-sourced its gateway at github.com/SpekoAI/gateway — an MIT-licensed single Go binary that runs as a sidecar and pins provider hosts with user-supplied keys, so BYOK mode does not communicate with Speko at all.
- Since starting the YC batch in late June, Speko's external usage has grown about 25 percent per week on average, front-loaded toward launch weeks; the hosted router and managed keys with consolidated billing are paid, while the gateway and BYOK setup will be free forever.
Why it matters: Speko turns the recurring R&D ritual of rebenchmarking every new speech model — which Bek repeated manually across four years of enterprise work — into an API call, so teams running thousands of voice calls a day no longer ship last quarter's models. External usage has grown about 25 percent per week since late June, and adoption hinges on whether enterprises trust its published rankings enough to route production traffic through a third-party.
Ask SkimNews
