Speko launched on Hacker News this week under the title "Launch HN: Speko (YC S26) – OpenRouter for Voice AI," pitching itself as a single API that benchmarks and routes across speech-to-text, LLM, and text-to-speech providers instead of locking you into one vendor's stack. I read the launch thread, went through Speko's docs, pricing page, and public benchmark leaderboard, and tried the router against my own account before writing any of this.
Short answer: Speko is a YC S26 startup that routes voice AI traffic across 20+ speech-to-text, LLM, and text-to-speech providers through one OpenAI-compatible API. It benchmarks models by language, latency, and cost, then picks the best combination automatically. Pricing runs +5% over provider rates on the Router plan, or a flat $0.09/minute on the bundled Infra plan, with $100 in free signup credit.
In my testing, the gap between the marketing pitch and what's actually documented was smaller than I expected for a week-old launch — the pricing page, docs, and benchmark site all agreed with each other, which isn't always true for a startup this early. Here's what Speko actually does, what it costs, and what's worth knowing before you point a production voice agent at it.
What you'll need
Speko is a developer tool, not a consumer app, so you'll want an existing voice pipeline or agent framework to point it at — LiveKit, Pipecat, or a custom setup already calling OpenAI-style speech and chat endpoints. Sign up at platform.speko.ai to get an API key and $100 in signup credit, which is enough to run real tests without a credit card commitment on day one. You'll also want to know your own priorities going in: are you optimizing for the lowest word error rate, the fastest response, or the cheapest bill per minute? Speko's routing is built around picking a winner for whichever of those you tell it matters, and it can't do that if you don't specify one. If you already have accounts with providers like ElevenLabs, Deepgram, or OpenAI's audio models, keep those keys handy too — Speko's BYOK gateway mode routes to your own provider accounts rather than reselling access.
Step-by-step: trying Speko's router
1. Create an account and grab an API key
Sign in at platform.speko.ai and generate a key. New accounts start with $100 in credit, so you can run a real workload before deciding whether to add a card.
2. Point your existing client at Speko instead of a single vendor
Speko speaks the OpenAI API, so if your app already calls OpenAI-style endpoints for transcription, chat, or speech synthesis, you swap the base URL and key rather than rewriting integration code. LiveKit users can drop it in as an AgentSession target; Pipecat users wire it in through the published OpenAPI spec at api.speko.ai.
3. Set a routing objective, not a specific model
Instead of hardcoding a model name, you request "auto" and tell Speko what you're optimizing for using request headers — Speko's docs point to X-Speko-Objective (latency, quality, cost, or balanced) and X-Speko-Max-Price as the levers that steer which provider handles the call.
4. Check the public benchmark leaderboard before you trust a claim
Before routing production traffic anywhere, I pulled up benchmarks.speko.ai directly. It lists 19 speech-to-text models scored on word error rate, latency, and cost — worth checking against your own target language, since Speko's own numbers show real spread between providers.
5. Decide between BYOK and hosted routing
If you want zero cloud dependency, Speko's gateway can run as a self-hosted, MIT-licensed sidecar that talks directly to your own provider accounts. If you'd rather not manage separate keys for every provider, the hosted Router or Infra plans handle key consolidation and billing for you, at a markup.
Example requests you can copy
These follow the pattern in Speko's own quickstart docs, which have you install the standard OpenAI SDK and swap in Speko's endpoint and key rather than learning a new client library:
- Transcription with automatic model selection: call the audio transcription endpoint with
model="auto"and let Speko pick the STT provider for the detected language. - Cost-capped routing: add
X-Speko-Max-Priceto a request so it never routes to a provider above your per-minute ceiling, even if that provider would otherwise win on accuracy. - Latency-first routing for live calls: set
X-Speko-Objective: latencyon real-time sessions, where a 66ms finalize time versus a 1.1-second one is the difference between a natural pause and an awkward gap. - Language-specific routing: Speko's own benchmark writeup notes that 4 different STT models win across 9 tested languages, so a request tagged with the right language hint avoids defaulting to whatever wins in English.
Common mistakes to avoid
The first mistake I'd flag: assuming "OpenRouter for voice AI" means Speko is free like browsing OpenRouter's model list is free. It isn't — the Router plan adds 5% on top of whatever the underlying provider charges, and that's before you've spent a cent on the models themselves. Second, treating every benchmark number as if it applies to every language. Speko's own leaderboard flags that 11 of the 23 models it tracks are only measured in English, so a model's headline word-error-rate might be unverified for the language your users actually speak. Third, skipping the objective headers entirely and letting "auto" mode guess — quality, latency, and cost routing pull in different directions, and not stating a priority means you get whichever default Speko ships with, not what you'd have picked yourself. Fourth, confusing the BYOK gateway with the hosted plans: the open-source sidecar is free to self-host indefinitely, per the founder's comments in the HN launch thread, but hosted routing and managed key consolidation are the parts Speko actually charges for. Fifth, not reading the benchmark methodology before trusting a leaderboard number for a compliance-sensitive use case — HIPAA and SOC 2 badges cover the platform, not a guarantee that any single routed model meets your specific regulatory bar.
What Speko costs
Speko's pricing is simpler than most routing platforms, with three tiers and no long-term commitment on the first two:
| Plan | Price | What it covers | Best for |
|---|---|---|---|
| Router | +5% on the provider's rate | STT, LLM, and TTS routing with measured selection and pre-response failover | Teams who already have provider accounts and just want smart routing on top |
| Speko Infra | $0.09/minute flat, all three services bundled | STT + LLM + TTS in one metered rate | Teams who'd rather not manage separate provider bills |
| Enterprise | Custom, rate locked for the contract term | Everything above plus deployment reviews and a dedicated SLA | Larger deployments needing a committed monthly minimum |
Every new account gets $100 in signup credit, and the self-hosted BYOK gateway is free indefinitely if you're willing to manage your own provider keys. Underlying model costs vary widely by provider — speech-to-text runs $0.0010–$0.0170 per minute of audio, and text-to-speech runs $10–$100 per million characters, per Speko’s pricing page, checked August 18, 2026.
The Hacker News launch: what people are actually debating
The founder, Beknazar Abdikamalov, pitched Speko as "OpenRouter for voice AI" and framed the core problem as teams picking a voice stack once and never revisiting it, even as better models ship monthly. That's not a cold start — per Speko’s YC company profile, he previously co-founded and was CTO at Hupo, a voice AI company that raised $14 million and counted Morgan Stanley, Prudential, HSBC, and Grab as clients, before spending what he describes as four years manually re-testing voice models by hand. Speko is that manual process turned into a product: a four-person, San Francisco-based team that's grown external usage roughly 25% per week since a soft launch in late June, ahead of this week's public HN post.
The pushback in the comments was worth reading past the pitch. One thread of criticism argued the industry is moving toward single end-to-end voice models, which would make Speko's cascaded STT-then-LLM-then-TTS routing less relevant over time — if one model handles the whole pipeline, there's less to route between. Founders and other commenters countered that most production teams still want to swap pieces independently for observability and control, and that cascaded architectures aren't going away as fast as end-to-end demos suggest. A second line of questioning was about on-device and local models eating into cloud routing's relevance; the response was that current production systems are still overwhelmingly cloud-based cascaded setups, whatever the demo reels show.
Tools that make this easier
If you're evaluating specific TTS or STT providers Speko routes to, my how to use ElevenLabs guide covers one of the vendors on Speko's leaderboard directly, including its own pricing outside of any router markup. For a comparison point on custom voice agents built the traditional way — a sales team configuring your stack instead of you wiring up an API — Leaping AI is worth reading before you decide whether to build or buy. If you just want to hear what a mainstream voice assistant sounds like without touching an API key, how to use ChatGPT Voice is the free, no-setup version of the same underlying idea. For how I vet claims like the ones in this piece before publishing, see how we test AI tools, and our AI tool ratings to compare this category against others. If you're new to wiring up AI provider accounts generally, kit for AI walks through what's actually worth setting up first.
My take
Speko's pitch holds up better than most week-old launches: the benchmark numbers, pricing, and docs all matched what I saw in the product, and the honesty about which models are English-only is the kind of detail a startup trying to look more finished than it is would probably hide. What gives me pause is the same thing HN commenters raised — routing across separate STT/LLM/TTS providers is a bet on the cascaded architecture staying dominant, and that's genuinely contested right now, not settled. If you're already running a cascaded voice stack and re-testing providers by hand every few months, Speko replaces that manual work for a 5% fee, which is a fair trade. If you're building fresh and end-to-end models already cover your use case, routing between three separate providers may be solving a problem you don't have yet.
Frequently Asked Questions
Is Speko free?
Not entirely. The self-hosted, MIT-licensed BYOK gateway is free indefinitely, but you still pay your underlying provider (OpenAI, ElevenLabs, Deepgram, and so on) for actual usage. Hosted routing adds a 5% markup on the Router plan or a flat $0.09/minute on the Infra plan, offset by $100 in free signup credit.
How long does it take to start using Speko?
About ten minutes if you already have a working voice pipeline. Sign up at platform.speko.ai, generate a key, and swap your existing OpenAI-compatible base URL for Speko's — most of the setup time goes into deciding your routing objective, not the integration itself.
What is the easiest way to try Speko well?
Start with the free $100 credit on a real (not toy) transcription or TTS job, set an explicit X-Speko-Objective instead of leaving it on default, and check benchmarks.speko.ai for your actual target language before trusting an English-language accuracy number.
Which voice AI providers does Speko route between?
Its public benchmarks list providers including OpenAI, Google, AssemblyAI, Deepgram, ElevenLabs, Cartesia, and several others across 19+ speech-to-text models alone, plus additional text-to-speech and speech-to-speech options — per Speko's own benchmark leaderboard.
Is Speko a good fit if I'm just testing voice AI for the first time?
Probably not yet. It's built for teams already running a voice stack who want to stop manually re-testing providers. If you're brand new to voice AI, a single free tool like ChatGPT Voice is a lower-stakes way to see what current models sound like before you need a router at all.