Speko is a voice model router that automates the choice of speech-to-text (STT), LLM, and text-to-speech (TTS) models for voice AI agents — based not on provider claims, but on benchmarks Speko continuously measures itself. Y Combinator’s official launch post frames it as “OpenRouter for Voice”: one API key and a single OpenAI-compatible endpoint give you access to 72 models spanning STT, LLM, TTS, and speech-to-speech (S2S). Accepted into Y Combinator’s Summer 2026 (S26) batch, Speko was founded by the former CTO of Hupo, which delivered voice AI to Morgan Stanley, HSBC, and Grab for four years. It started out focused purely on automatic voice model routing, then added a Managed Agents Platform for running reusable agents across both browser and phone, plus Gateway — an open-source local runtime (early preview) that streams directly to providers. The result is a stack where developers can pick their own entry point.
Key Features
- Automatic routing based on per-language benchmarks: 17 STT models, 23 LLMs, 22 TTS models, and 10 S2S models — 72 in total — are continuously benchmarked across 10 languages including English, Spanish, Arabic, French, German, Hindi, Norwegian, Tamil, Telugu, and Filipino. Requests are routed to the best model according to
optimizeFor(balanced/accuracy/latency/cost), which sets how accuracy (WER and similar), latency, and cost are weighted. All measurements are published publicly at benchmarks.speko.ai - A single OpenAI-compatible endpoint: One base URL and one API key cover STT, LLM, and TTS. Existing frameworks like LiveKit’s OpenAI plugin or Pipecat work without code changes — just point baseURL at Speko
- Automatic pre-response failover: If the selected provider fails, Speko automatically retries with the next-ranked candidate from the benchmarks. Responses carry provider, model, and failoverCount fields plus headers like X-Speko-Provider, so you can trace what actually ran
- TypeScript (LiveKit) / Python (Pipecat) SDKs and an official MCP server: Alongside typed SDKs and a browser WebRTC client, Speko officially ships an MCP server (
https://mcp.speko.ai/mcp) that lets coding agents such as Claude Code, Codex, and Cursor create, deploy, test, and migrate agents - Managed Agents Platform and Gateway (early preview): Beyond the routing API (Relay), there’s a management platform for reusable agents — each with its own prompt, voice, and tools — running across browser conversations and inbound/outbound phone calls, plus Gateway, an open-source local sidecar that connects directly to providers while keeping BYOK credentials inside your own process
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Router | Provider list rate + 5% | Usage-based pricing with automatic selection and failover included. No commitment, email support |
| Speko infra | $0.09/min (all-inclusive) | STT, LLM, and TTS bundled into a flat per-minute rate. No commitment |
| Enterprise | Contact sales (contract-locked) | Deployment review, dedicated support, contractual SLA, monthly minimum spend |
Pricing reflects August 2026. New sign-ups receive $100 in signup credits. Both Router and Speko infra are currently in Public Preview (no SLA). Check the official site for current pricing.
Pros and Cons
✅ Pros
- Accuracy, latency, and cost measurements for every model are published openly at benchmarks.speko.ai, so you choose models from real numbers rather than vendor self-reporting
- One API key and an OpenAI-compatible endpoint let you switch across nearly 30 providers and 72 models — OpenAI, ElevenLabs, Deepgram, Cartesia, Hume, and more — without absorbing each contract and SDK difference yourself
- Automatic pre-response failover means a single provider’s outage is less likely to take down your entire voice pipeline
- The MCP server lets you drive agent creation, deployment, and testing on Speko conversationally from coding agents like Claude Code
- $100 in signup credits makes both Router and Speko infra easy to try with real requests
⚠️ Cons
- Router and Speko infra are both Public Preview with no SLA; production availability guarantees require an Enterprise contract
- A young service, freshly accepted into the YC S26 batch, with limited track record, case studies, and community knowledge so far
- Telephony connectivity is out of scope — SIP and phone carrier integration must be arranged separately on the application side
- Unlike Vapi or Retell AI, Speko doesn’t manage the conversation loop for you; it assumes you operate your own runtime such as LiveKit or Pipecat
Comparison with Similar Services
| Criteria | Speko | Vapi | Retell AI |
|---|---|---|---|
| Positioning | Voice model router (the STT/LLM/TTS selection layer) | Hosted development platform that runs the whole conversation loop | Hosted platform specialized in phone voice agents |
| Conversation loop | Held by your app (LiveKit/Pipecat, etc.) | Orchestrated by Vapi | Orchestrated by Retell |
| Telephony | Out of scope (your app’s responsibility) | Connected, billed separately | Retell’s carrier service or your own SIP, billed separately |
| Model selection | Automatic, benchmark-driven — or pinned | Developer configures supported STT/LLM/TTS | Chosen from Retell’s supported models and infrastructure options |
| Public benchmarks | All models published at benchmarks.speko.ai | Not published | Not published |
| Pricing model | Usage-based (Router: provider cost +5% / Speko infra: $0.09/min) | Usage-based hosting and orchestration fees plus per-model/per-call charges | Component-based (voice infrastructure plus TTS/LLM/telephony, etc.) |
Speko publishes its own comparison against both services on its official site (confirmed as of July 31, 2026), emphasizing the positional difference: Vapi and Retell AI take on the conversation loop and operations, while Speko handles only STT/LLM/TTS selection.
Who It’s For
- Developers who already run their own voice agent runtime on LiveKit or Pipecat and want to externalize just STT/LLM/TTS selection and switching
- Multilingual voice product teams tired of researching the best voice model for each supported language, over and over
- Teams looking to reduce the risk that a single provider’s outage stops their voice pipeline outright
- Anyone who wants to hand voice agent building and deployment to coding agents like Claude Code or Cursor
Summary
Speko is a routing API — one YC itself calls “OpenRouter for Voice” — that automates STT, LLM, and TTS selection for voice AI based on public benchmarks rather than vendor claims. The key difference from Vapi and Retell AI is that Speko doesn’t manage the conversation loop itself; it specializes in the model selection layer underneath. The payoff is largest for teams that already run their own runtime on LiveKit or Pipecat. Since it only launched in mid-2026 and remains in Public Preview, be mindful that the SLA and track record production deployments usually depend on aren’t there yet.