PolyAI opens voice-agent platform, claims Raven 3.5 beats GPT-5 on 4 benchmarks
PolyAI said it is opening its Agentic Dialog Platform to all enterprise builders, widening access to software the company says has resolved more than 1 billion customer conversations for clients including FedEx, Unicredit, PG&E, Marriott and Foot Locker. The launch includes a natural-language Poly Agent Builder, a code-first Agent Development Kit for building and deploying agents from a developer’s IDE and terminal, and PolyPhone, which turns website content into a voice agent in about 10 minutes.
PolyAI’s research team also published a paper saying its specialist Raven 3.5 model beat GPT-5 and Claude Sonnet 4.6 on four customer-service benchmarks while staying under 300 milliseconds of latency. The company said the model was trained on real customer conversations for live phone support and evaluated on English quality, response style, date-time logic and multilingual quality across 23 languages.
From the sources (5 posts)
@rohanpaul_aiVoice AI might be the biggest productivity boost you can add to almost any office job. And with PolyAI’s Agentic Dialog Platform now open to every enterprise builder, Voice AI has gone from a 6-figure annual contract to a free trial you ca
@archiexzzzwhat i’ve observed is that edge cases in voice are insanely high. conversational drift and "hacky" engineering make things more complex. raven is a model trained end-to-end on 1 billion real enterprise conversations. the “agent harness” (r
@rohanpaul_aiCan a smaller model purpose-built for one domain beat a frontier general model that's 100× its size? A recent paper showed yes — and not by a small margin. Raven 3.5 from PolyAI shows that a smaller specialist model can beat bigger gener
@rohanpaul_ai🧵 2. The problem is that general LLMs can sound smart, but live phone support needs fast replies, careful instruction following, natural speech, and reliable handling of messy caller requests. PolyAI’s answer is not a new giant model, but
@rohanpaul_ai🧵 4. PolyAI’s bet is that voice agents should be built for phone calls from day 1, not copied over from chatbot logic. That means the agent needs to handle interruptions, messy phrasing, accents, silence, back-and-forth corrections, and th