No Model Picker, No API Keys: Why the Model Set Is Curated

What we are not selling

There is no model marketplace here. No dropdown of hundreds of models, no leaderboard, and no bring-your-own-keys — you cannot plug in your own OpenAI, Anthropic, speech-recognition or speech-synthesis account. The one bring-your-own thing on this platform is a phone number.

The models are curated and platform-managed: a small vetted set, picked for latency and answer quality, replaced when something better ships. This post argues that trade honestly, including the half you lose.

What you give up

Model choice, and the things that come with it.

• You cannot pin a version. If your compliance process requires a named model at a frozen version for the next eighteen months, this is not the platform for that.
• You cannot self-host. If the requirement is that weights run on hardware you own, no managed platform solves it, including this one.
• You cannot chase a benchmark. A model tops a leaderboard on Tuesday and you cannot switch to it on Wednesday. We will get there when we have run it on real conversations.

If any of those three is a hard requirement, stop reading — the rest of this will not change your mind, and you should not find that out after signing up.

What you get

One price, and it contains everything

Voice is from ₹2 a minute — ₹2 for standard voices, ₹2.5 for premium ones. That rate is all-inclusive: the telephony, the speech recognition, the language model and the speech synthesis are all inside it. No per-language surcharge. No provider passthrough. No platform fee plus costs. And it is billed in 30-second pulses rather than rounded up to the whole minute, so four short calls cost roughly what four short calls should.

That is only possible because we control which models run. A platform that lets every tenant choose freely cannot quote a flat inclusive rate; it has to charge you a platform fee and pass the model bill through, which is why those platforms publish a per-minute number that turns out not to be the number.

Nothing to reconcile

The version of this product where you bring your own keys means four vendor relationships, four invoices in three currencies, four sets of rate limits, and a month-end where somebody works out which of them caused the spike. Chat replies here are usage-based on one bill, in rupees, through Razorpay and UPI. There is one number to look at.

Nothing to keep up with

Models are replaced as better ones ship, and your agent does not change shape when that happens — the prompt, the knowledge base and the tools are yours and carry over. The work of reading release notes, testing a new model on Indic speech and deciding whether it is actually better is work we would rather do once than have every customer do separately.

How the voice stack is actually built

The pipeline is cascaded: speech to text, then a language model, then text to speech — Sarvam for recognition, Cartesia for synthesis. Every stage is a separate model, which is why the same knowledge base and the same instructions serve a phone call and a WhatsApp thread.

What that rules out, and we would rather say it than let you assume otherwise: there is no speech-to-speech model and no realtime voice model on offer here. Not Gemini Live, not a realtime API from anyone. Those are a genuinely different architecture with genuinely different tradeoffs, and if that is specifically what you are shopping for, you are not shopping for this.

We also do not publish a latency figure. Carrier-call performance is not something we have measured to a standard we would stand behind, and a number we cannot defend is worse than none.

Free tier and paid tier

The free tier runs on Prana LLM and gpt-4o-mini. The premium language models sit on the paid tiers. That is the whole model story: two shelves, not a catalogue.

The trial gives you 25 voice minutes and 200 chat replies with no card. After that, Pay As You Go has no monthly fee, Scale is ₹4,999 a month or ₹54,989 a year, and Enterprise is negotiated. Prices are pre-GST.

Languages come from the platform, not the model picker

The agent replies in 37 languages, including all 22 scheduled Indian languages in their native scripts, and handles Hinglish code-switching. One knowledge base serves all of them — you do not maintain a Marathi copy of your refund policy, and there is no per-language surcharge to work out before you enable one.

On a marketplace platform, "which model is best for Tamil" becomes a question you have to answer, re-answer every quarter, and pay differently for. Here it is a setting.

The honest summary

A curated model set is a smaller promise than a marketplace, and smaller promises are easier to keep. You are trading the right to pick a model for one predictable rupee figure, a single invoice, and someone else's job to notice when a better model ships. For most businesses running customer conversations, that is the better half of the trade. For a team with a specific model as a hard requirement, it is not — and we would rather you know that on this page than three weeks in.

Get started

Create an agent, give it your website to crawl, and have a browser voice call with it before a phone number is involved. That tells you more about the model set than any specification sheet.

Try it free →

No credit card required • Trial includes 25 voice minutes and 200 chat replies