Tazo is an audio decision model for teams shipping live voice agents. You pass raw call audio, a text prompt, and the options you want scored. It returns a calibrated probability for each option, so the agent can choose among respond, wait, yield, backchannel, escalate, or an early tool call without waiting for a transcript. The API is live with this launch, priced at $0.20 per hour of audio, and new sign-ups get free credits. The failure this launch targets shows up after the words are already correct. In a typical cascaded stack, the choice to speak, wait, yield, or escalate still depends on a transcript and a silence timer. That path drops pitch, pauses, and tone, which often show whether a caller has finished or is still thinking. Tazo runs beside the existing speech-to-text, language model, and text-to-speech stack and scores options you define on each request, so a new decision is a new list rather than a new model. The company reports 70.9 percent versus 46.4 percent for speech-to-text plus Jev on real production calls, latency under 200 milliseconds, a second-place rank on the MMAU audio intelligence benchmark, and a none-of-the-above option for out-of-scope or unsure states. It can also start a tool lookup while the caller is still talking, so the result is ready when the turn ends. Tazo Labs is a Bengaluru company founded in 2026. Naman Jain, an AI researcher who studied computer science at IIT Gandhinagar, announced the product. The default API model is tazo-large.