The company, founded by Stanford AI Lab and Y Combinator alumni, claims 99.4% on the Big Bench Audio benchmark, the highest of any real-time model, and offers cost-plus pricing from 2 cents a minute.
SAN FRANCISCO, Nov. 10, 2026 — ThunderPhone, a startup whose platform runs AI agents on phone and web calls, launched its v3 architecture and an agent-building system called Zeus on Monday.
While typical voice AI systems power calls by connecting speech-to-text, language and text-to-speech AI models in sequence, the ThunderPhone v3 architecture instead runs many models on each call, including several transcription models and language models concurrently, said CEO Alex Kolchinski. The company has retired its fixed product tiers in favor of a custom model mix per customer use case.
Zeus, which the company calls an agentic harness for voice, automates the model tuning ThunderPhone engineers previously did by hand for each enterprise account, Kolchinski said.
“Zeus works like a coding agent for voice. It builds voice agents and constantly improves their performance and drives down their cost based on what it sees on simulated and real calls,” Kolchinski said.
The tool builds agents from recordings, call center scripts or imports from rivals Vapi, Retell, Bland and ElevenLabs. On simulated and live calls, it finds errors, tests fixes and backtests changes against past calls within user-set spending caps.
ThunderPhone reported a 99.4% score on the Big Bench Audio reasoning benchmark, posting its evaluation set at huggingface.co/datasets/kolchinski/bba-storm-eval. Only StepFun’s StepAudio 3 Realtime has a higher published score at 99.7%, but ThunderPhone said that model is too slow for live phone calls, which the company says makes v3 the highest-scoring model that can handle them.
ThunderPhone has also introduced cost-plus pricing, charging customers underlying vendor rates plus a published 17% to 25% margin, which the company claims makes it the cheapest voice AI platform on the market. Kolchinski said calls start at 2 cents per minute, though 5 cents is common for more complex agents and complex use cases can run 10 to 15 cents per minute. ThunderPhone also supports custom pricing under 1 cent per minute for high-volume enterprise use cases.
“To our knowledge, nobody else in voice AI undercuts us on price or publishes their margin,” Kolchinski said.
Kolchinski and Alec Bell, ThunderPhone’s CTO, founded ThunderPhone (Autophonix, LLC) in 2025. The bootstrapped startup says it is profitable and growing, though without disclosing precise revenue. Kolchinski previously co-founded Y Combinator W21 startup Mezli, left a PhD in the Stanford AI Lab, and was a Google associate product manager. The launch follows Vapi’s $50 million Series B in May and Deepgram’s October deal to buy Pipecat maker Daily.
The platform is live and free to try, with details at thunderphone.com/blog/introducing-zeus.
About ThunderPhone
ThunderPhone runs voice AI agents in 62 languages for uses including receptionists, customer support, appointment booking, outbound sales and phone interpreting. The platform is HIPAA and GDPR compliant and works through a web app, API, MCP server, LiveKit, Pipecat, Zapier and n8n at thunderphone.com.
Media Contact
Alex Kolchinski, CEO
[email protected]
(510) 858-4272
SOURCE ThunderPhone
