Open-weight models on Open Voice Shield: OpenRouter and Groq

OpenRouter for analysis. OpenRouter fronts dozens of hosts running open-weight models behind one API. The analysis step, the part that reads a transcript and writes the verdict, can now run on gpt-oss 120B, Llama 3.3 70B, Mistral Small 3.2, Qwen3 235B, DeepSeek V3.2, Gemma 3 and others. We route by throughput rather than by price, because the cheapest host of a 120-billion-parameter model can take most of a minute to answer and a fast one takes a second. The difference in cost is a fraction of a cent. Each call records what OpenRouter charged, not a list price.

Groq for transcription. Groq runs the same Whisper large-v3 weights the platform already knew, on hardware that returns a minute of audio in about a second, and it answers with timed segments, so the caller and the callee stay interleaved in the transcript the way the conversation went. It is now an STT provider next to OpenAI and the local Whisper on the worker.

Both are chosen per AI tier in the admin settings, the same place the tier's fee is set. Nothing changes for a switch or a phone system sending calls through the hop.

Until now the platform could use frontier models from commercial labs, or a Whisper on the worker for transcription, and nothing in between. The middle is where most of the volume wants to be: models whose weights are public, run by hosts that compete on price and speed. Every tier can now be built from whichever class fits it.

The cost of reading a call fell, and the per-call analysis fee followed on every tier: Economy $0.015, Power $0.02, Max $0.05 a call, transcription and analysis included. A three-minute call on the default plan and tier is $0.08 all in. Per-minute rates did not change.

The verdict keeps its shape: a probability, a category, the entity the caller claimed to be, numbered red flags quoting the call, a TCPA note and a block, review or allow recommendation. The default tier now answers in about a second once the transcript is ready, which shortens the wait between a hangup and an alert. If you want the deepest written analysis on every call, Max still runs a frontier model. If you want to sample everything, use Economy. Switch tiers on the dashboard. The change applies to the next call.

Your audio goes to one transcription provider and your transcript to one analysis provider, the ones behind your tier, and each call record names both and what they cost. That has not changed. There are more providers to choose from behind each tier.

An on-premises licence puts the platform on hardware you operate. Every provider is a key you hold, every tier is a pair you set, and the admin settings now show what each model costs beside what you charge, so the margin of a tier is a number on the page rather than a guess.

• Keep audio at home. Run Whisper on your own worker for transcription and only the text leaves for analysis, or run nothing outside at all if your analysis model is on your side of the wire.
• Spend where it matters. An open-weight model through OpenRouter for the bulk of traffic, a frontier model for the routes or clients that justify it, the same verdict format across both.
• Measure before you switch. The comparison harness that ranked the models for us ships with the platform: point it at your own labelled calls and read agreement, cost and latency per model. Any stored call can be re-analyzed with another model from the admin side, so a decision is made on your calls, not ours.
• Price it your way. Set the fee per tier and the per-minute rate to your market. The cost side is measured per call, and OpenRouter reports its real charge.

We ran a labelled set of scam and legitimate transcripts, plus real calls from the platform, through each candidate and compared verdict, category and probability with the frontier models we had been using, then had a separate model score every written analysis blind for how well its red flags quoted the call. gpt-oss 120B agreed with the frontier verdicts on the real calls and answered in about a second through OpenRouter's fastest hosts. Several smaller open models were as close on the verdict but weaker in the written evidence, and a few were slow enough to rule out. Groq's Whisper matched the transcription accuracy we had on the same recordings at a fraction of the price.

Two limits. The labelled set is small and synthetic, so it ranks models. It is not an accuracy figure and we do not publish one. And routing by throughput means the host, and therefore the exact cost, can vary from call to call, which is why each call records the actual charge rather than a list price.

Details of the models and fees are on the AI modes page. The release notes are on what's new. New here? Start free with $5 credit.