Speko Review (2026): Pricing, Features & Honest Verdict

Speko is a routing layer for voice AI that publishes measured benchmarks for 62 speech and language models and charges either 5 percent on top of provider rates or $0.09 a minute all-in. The benchmark data is the real product. Best for: teams building voice agents. Price: From $0.09/min or +5% on provider rates (free plan: no). Rating: 7.5/10

What is Speko?

Speko launched on Product Hunt on 27 August 2026 at #5 for the day, describing itself as OpenRouter for voice. It is operated by Speko Group, Inc. from the United States and is a Y Combinator company. The pitch is that nobody knows which speech model is best for their language, because the vendors all benchmark in English and market on the result, so Speko measures them itself and routes your traffic to whichever one wins for the job in front of it.

There are two surfaces. Speko Router is the hosted, provider-neutral data plane at router.speko.dev, with typed operations for model discovery, transcription, synthesis and LLM responses. You either let it pick a route by latency, quality, cost or a balanced objective, or you pin a specific provider and model, and every response reports which route it took in structured metadata. Speko Gateway is a separate open customer-side runtime that sits beside LiveKit or Pipecat, keeps your provider credentials inside your own process, and exposes one local streaming protocol across every provider. A third hosted Agents API at api.speko.dev owns agents, sessions, phone calls and telephony.

What Are Speko’s Key Features?

A benchmark board that admits what it has not measured

This is why the product is worth your time. Speko publishes a catalogue of 62 models from about 30 providers, split into 17 speech-to-text, 13 LLM, 22 text-to-speech and 10 speech-to-speech entries. For every STT model it lists batch word error rate, streaming word error rate, p50 finalize latency and cost per minute of audio. AssemblyAI’s Universal-3.5 Pro posts 2.0 percent WER at 66ms and $0.0075 a minute. Deepgram’s Nova-3 posts 9.8 percent at 106ms and $0.0048. Modulate’s Velma 2 costs $0.0010 a minute and posts 4.4 percent. Where a number does not exist, the cell says “not measured” rather than being filled in. The page states plainly that nothing on it is estimated to fill a column.

Language coverage as the actual argument

Speko’s own headline finding is that 11 of 23 STT models are measured in English only, including the model that tops the English table, so their rank in any other language is unknown. Across the nine languages it does measure, four different models win. If you are shipping a voice agent in Arabic, Hindi, Tamil, Telugu or Norwegian, that is the single most useful thing anyone has published for you, and it is free to read without an account.

Routing with pre-response failover

The Router does measured selection and pre-response failover, meaning it can switch providers before the caller hears a problem rather than after. You give it an objective, latency or quality or cost or balanced, and it picks from eligible routes. The LLM board carries its own metrics that are unusual and useful: task completion rate on graded tasks, time to first token, dead-air percentage of turns, and a fabrication rate for off-policy responses. Claude Haiku 4.5 posts 2 percent dead air against gpt-5.6-luna’s 70 percent, which is exactly the kind of thing that ruins a voice agent and never appears on a model card.

Gateway keeps your keys in your own process

The open Gateway is the option for teams that own their conversation loop. It runs beside LiveKit or Pipecat, holds BYOK provider credentials locally, and never hands them to Speko. Its public repository also owns the canonical wire contract the hosted Router consumes, so the two stay compatible. That is a real answer to the obvious objection about putting a middleman between you and your audio.

How Much Does Speko Cost?

Speko publishes two self-serve paths and one contract. Router charges 5 percent on top of whatever your provider’s published rate is, and you keep paying the provider’s rate for speech-to-text, LLM and text-to-speech. Speko infra is $0.09 a minute all-in, with all three model legs included. Enterprise is custom, locked for the term, and requires a committed monthly minimum that is never quantified. Every account gets $100 in signup credit, which buys 1,111 minutes of audio at the flat rate, or roughly 18 and a half hours.

Do the division and the flat rate looks poor. For $0.09 a minute to beat the Router path, your raw model spend has to exceed $0.0857 a minute, because Router only adds 5 percent on top. Retell AI publishes a worked breakdown of a typical voice minute at $0.04 for the LLM plus $0.015 for TTS, and Speko’s own board prices decent STT between $0.0010 and $0.0170. Add those and you are near $0.061 a minute in raw model cost, which the Router turns into about $0.064. The flat $0.09 rate is roughly 40 percent more expensive for that stack, and nothing on the pricing page points this out.

Against the field, Router is the interesting number. Vapi charges a $0.05 per minute platform fee and passes model costs through at cost, or at zero if you bring your own key. Speko’s Router charges no per-minute fee and takes 5 percent of model spend instead. For Vapi’s fixed $0.05 to beat a 5 percent cut, your model spend would need to exceed $1.00 a minute, which is more than fifty times any rate on Speko’s own board. Retell AI starts at $0 with $10 in free credits and runs $0.07 to $0.31 a minute for voice agents. On pure economics for a self-assembled stack, Speko Router undercuts both by a wide margin.

The catch is written on the page in small print. Both self-serve tiers say “Public preview” under Access and “None during preview” under SLA. There is no uptime commitment at any price you can buy with a card. Speko is asking you to put a routing hop in front of your production voice traffic and offering no guarantee it stays up.

Who is Speko Best For?

Use Speko if you are building voice agents in a language other than English. The benchmark board alone will save you a month of evaluation work, and the routing gets you a model that was actually measured on your language instead of one that markets well in English.

Use Speko Router if you already pay providers directly and want failover without a per-minute platform fee. Five percent on your existing spend is the cheapest managed routing on offer right now, and the Gateway option means you can keep your keys.

Skip Speko if you need an SLA today. Preview access with no uptime commitment is fine for a prototype and wrong for a product taking customer calls, and Enterprise is the only tier with a contractual SLA. Skip the $0.09 flat rate specifically unless your model stack is unusually expensive, because for most stacks the Router path costs less for the same service. And skip it if your security review needs a published security page, because /security returns a 404 and only terms and privacy exist.

Best Speko Alternatives

Vapi is the incumbent for hosted voice agents. It charges $0.05 a minute as a platform fee and passes STT, LLM and TTS through at cost, dropping to zero if you bring your own API keys, with 10 call concurrency included and $10 per additional line per month. HIPAA is a $2,000 a month add-on and Zero Data Retention is $1,000 a month. Vapi is the more complete platform. It is also the more expensive one at any realistic model spend.

Retell AI starts at $0 with $10 in free credits and no commitment, running $0.07 to $0.31 a minute for voice agents and $0.002 or more per message for chat. Its pricing page publishes a component breakdown, $0.04 LLM plus $0.055 voice infrastructure plus $0.015 TTS for a $0.11 minute, which is more transparent than most. Twenty concurrent calls are included. Pick Retell if you want the agent platform rather than the routing layer.

Going provider-direct is a real alternative and Speko’s own catalogue makes the case. Deepgram Nova-3 is $0.0048 a minute, Inworld’s Realtime STT-1 is $0.0025 at 139ms, Soniox stt-rt-v5 is $0.0020 at 78ms. If you serve one language and one provider covers it well, you do not need a router, and you save even the 5 percent. Speko’s board is still the best free tool for deciding which provider that is.

Final Verdict: Is Speko Worth It?

The benchmarks are the product and they are excellent. Publishing streaming and batch word error rates, finalize latency, cost per minute and language coverage for 62 models, then labelling the gaps “not measured” instead of filling them with vendor claims, is more useful than the routing it exists to sell. The finding that 11 of 23 speech models have never been measured outside English, and that four different models win across nine languages, is a real contribution to a category that runs on marketing numbers. The Router pricing is honest too, and at 5 percent it undercuts Vapi’s fixed per-minute fee by a factor that only reverses if you spend more than a dollar a minute on models.

Two things stop this being an 8. The $0.09 flat rate is the worse of the two self-serve options for almost any stack you would actually build, and a company this careful with its benchmark methodology should say so rather than list them side by side as equals. And the whole self-serve product is a public preview with no SLA, on infrastructure that sits in the path of live customer audio. Use the benchmark board today, whatever you end up buying. Put Router in front of a prototype and see how the failover behaves. Wait for a general availability announcement and an uptime commitment before you route paying customers through it.