Speko Review

61 models benchmarked across 10 languages, published at benchmarks.speko.ai. That's a specific, checkable claim, not a marketing adjective. Most voice infra vendors bury routing logic; Speko exposes it via response headers like Speko-Provider and Speko-Model, so you can audit what actually happened.

Two pricing paths: $0.09/min all-in on their infrastructure, or provider rate plus 5% through the Router. Both come with $10 signup credit and no tiers to negotiate. That's clean, but it's also usage-based with no published enterprise floor, so cost at scale is a spreadsheet exercise, not a lookup.

The open Gateway with BYOK credential custody is the smart hedge here. It lets LiveKit and Pipecat shops keep sensitive keys local while still using Speko's routing brain elsewhere. Lock-in risk is lower than a typical managed router because of that split.

Sixty-one models across 10 languages, published at benchmarks.speko.ai. That's a real design decision: most voice tooling still treats English as the default and everything else as a rounding error. Building routing around per-language performance rather than one benchmark leaderboard shows someone on this team understands that voice craft isn't uniform across markets.

But this is systems design, not creative tooling. There's no mention of prosody control, voice persona management, or brand-voice consistency across providers — the object model is provider/model/region, not tone/character/emotion. If your job is shaping how an assistant sounds and feels, Speko sits one layer below where you work; it decides which engine answers the phone, not what the voice says or how it says it.

The RFC 9745 deprecation policy and 90-day sunset notice tell me this team thinks in years, not sprints. That's the right posture for something your agents depend on daily. The tradeoff: you're adopting a routing dependency to solve a problem — provider fragmentation — that only exists because the voice AI market itself is still consolidating.

Two SKUs. Speko Infrastructure at $0.09/min flat, covering STT, LLM, TTS combined. Routed Provider Pricing at provider rate plus 5%. Both visible, both usage-based, no seats.

Run the math at scale. A 50-agent voice desk doing 20 min/day/agent, 22 workdays: 22,000 minutes/month. At $0.09 flat that's $1,980/month, ~$71K over 3 years before growth. Provider-routed cost floats with whatever the underlying vendor charges — harder to lock into a budget line. $10 signup credit gets you started, not far.

No enterprise tier posted, so high-volume discounting needs a direct conversation — fine for procurement, less fine for finance modeling year 3 without a call. Usage-based billing means no flat invoice to forecast against; good for pay-as-you-grow, bad for anyone who needs a fixed number for next year's budget. No contract terms, renewal windows, or termination language visible on the pricing page itself.

I mix voice for conversational agents, and the failure mode I dread is a provider swap mid-session that changes prosody without telling anyone. Speko's response headers name the exact route taken per request, which means when a take sounds wrong I can actually trace it instead of guessing. That's a real win for post-mortem work.

The idempotency requirement on every POST and WebSocket upgrade is the kind of detail that only shows up when you've been burned by a duplicate TTS render eating budget on a retry. Good instinct, small tax on setup.

What's missing for my chair: nothing in the routing logic touches performance direction. The benchmarks at 61 models across 10 languages tell me which model wins on accuracy, not which one sounds right for a brand voice or a character read. Objectives are balanced, quality, latency, cost — none of them are 'sounds like this reference clip.' Fine for IVR and support bots. For anything with a performance to protect, I'm still doing that judgment call by ear, outside the router entirely.

There's no dashboard to fall in love with here, and that's kind of the point. You're integrating an HTTP and WebSocket layer, reading response headers like Speko-Provider and Speko-Model, and trusting a routing decision most days without looking at it twice. The 90-day deprecation notice and RFC 9745-style headers tell me someone's thought about what happens when things change under you, which matters more than any onboarding screen ever will for a tool like this.

What worries me a little is the Idempotency-Key requirement on every POST — smart engineering, but it means your first hour is reading docs, not shipping audio. That's homework before the fun part.

The split between the hosted Router, the open Gateway, and a separate Agents API means three mental models to hold at once, and I couldn't find much on how errors actually surface in practice beyond the documented envelope. For a plumbing product, that's the real test: not the demo, but whether provider_unavailable shows up gracefully three months in.

Middleman businesses live or die on one question: do you trust the middleman's judgment more than your own? Speko's whole pitch is that English-only benchmarks mislead you, so they built one across 10 languages and 61 models. Fine. But the benchmark lives at benchmarks.speko.ai — their own domain, their own methodology, no third-party audit mentioned anywhere I found.

I've seen this shape before: infrastructure vendor publishes the metric that just happens to justify their product. Doesn't mean it's wrong. Means I'd want an independent eval before trusting auto-mode with quality objectives on production traffic.

Pricing is at least legible: $0.09/min all-in, or provider-rate plus 5% if routed. No enterprise tier visible, which is normal for a usage-based tool this size. The Gateway's BYOK design is the honest exit hatch here — if the Router's judgment goes sideways, you're not fully locked to their infrastructure.