Is Prism state of the art?Pending benchmarks.Announced Oct 28, 2026; no independent results yet.
Rime's speech-to-speech model for voice agents on phone calls. It works from the caller's audio, writes each reply as text before speaking it so guardrails can check every word, and runs tool calls while it talks. English only, four voices, over a WebSocket API at $0.08 per minute; enterprises can deploy it in a VPC or on-prem. No weights.
The details so far
- Lab
- Rime
- Announced
- Oct 28, 2026
- API
- Rime SDK (Python, Node.js, Go, Rust) or WebSocket wss://prism.api.rime.ai/v1/realtime in US East; LiveKit and Pipecat integrations
- Open weights
- Closed · None
- Task
- Speech-to-speech voice agent model: one WebSocket connection replaces speech-to-text, a language model and text-to-speech
- Text layer
- Each reply is generated as text before speech and sent with the audio, so rules and disclosures can check it first
- Language
- English only
- Voices
- aurora, wawona, brando and astra (Prism's own voices; Rime's text-to-speech voices do not work with it)
- Audio
- Headerless PCM16 mono, 16 kHz in and 24 kHz out; phone audio needs a bridge
- Context
- 65,536 tokens for instructions, tools, history and tool results
- Pricing
- $0.080 per minute on Flex, $0.064 on Scale ($100/month); $50 of free usage each month
- Deployment
- Rime cloud on self-serve plans; cloud, VPC or on-prem on Enterprise, with HIPAA BAA and SOC 2 Type II
- Fine-tuning
- Rime trains Prism on a customer's calls; self-serve RL and SFT APIs are being built with beta customers
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 56.6 on Artificial Analysis's τ-Voice, run by Rime on that harness with no reasoning step, ahead of OpenAI and Google realtime models that do not reason first by 17 to 30 points; 82.5 in the telecom domain vs 45.6 for GPT-Live-1.
Lab claim: 0.57 s to first audio after the caller's turn ends on Artificial Analysis's speech-to-speech benchmark; Rime's blog and CEO also describe about 300 ms response to user audio.
Lab claim: a Prism trained on 10 hours of Revmo's calls placed 94% of orders only after the caller confirmed, vs 57% for Revmo's OpenAI agent.
Sources
LABRime blog: Introducing Prismrime.ai/resources/prism-first-speech-to-speech-for-enterprise · Oct 28, 2026LABRime docs: Prism overviewdocs.rime.ai/docs/prism/overviewLABRime pricingrime.ai/pricingLABLily Clifford (Rime) on X: Prism launchx.com/lilyjclifford/status/2115521147890520503 · Oct 28, 2026Tracking continues until an independent board lists it.