ThunderPhone

ThunderPhone builds AI agents that answer and place real phone calls — inbound support, outbound campaigns, or a voice agent embedded in a web app. Its pitch against the incumbents is blunt and mostly about unit economics: 2¢ and up per minute, where it lists Vapi, Retell and Bland at 7¢ and up. Surcharges apply for long prompts, added intelligence and premium voices.

Accuracy architecture. Rather than relying on one transcription model, ThunderPhone fuses multiple transcripts and combines them with direct audio-to-LLM processing. A dual-model setup — a fast model alongside a thinking model — is what keeps agents on script without making them sound robotic when a caller interrupts. The company reports 99.4% on the Big Bench Audio benchmark and a first-place ranking on it as of July 2026.

Conversation handling. Interruptions and backchannel responses ("mm-hm", "right") are handled natively, and agents support 40+ languages including switching language mid-call. Knowledge grounding comes from uploaded documents with retrieval, so answers cite your material rather than the model's priors.

Operations. Live supervision lets a human watch a call in progress, with AI-guided handoff when the agent should step aside. Native SIP trunking, DTMF capture and voicemail detection cover the telephony layer; A/B testing and closed-loop monitoring cover iteration. HIPAA and GDPR compliance is available with BAA and DPA.

Who it suits. Teams running call volume where per-minute cost compounds — appointment scheduling, qualification, support triage — and regulated industries that need a BAA in place. Built by Stanford AI Lab and Y Combinator alumni; CEO Alex Kolchinski. The operating entity is Autophonix, LLC d/b/a ThunderPhone.