Rime launched Prism on October 28, 2026, a speech-to-speech model aimed at enterprise voice agents. Its pitch is control: Prism writes what it is about to say as text, so rules can read, block or rewrite the reply before it is spoken, while the model still hears the caller. Secondary coverage, though, repeats a "300ms, 4x faster than GPT-Live" claim that does not appear in Rime's own launch post. This post lays out the claim next to what Rime actually published.
TL;DR: claimed vs published
| Item | Secondary coverage | Rime's own post |
|---|---|---|
| Response time | 300ms, "4x faster than GPT-Live" | About 500ms; 0.57 seconds to first audio, about half the wait of GPT-Live-1, GPT-Realtime-2.1 and Gemini 3.8 Live |
| Benchmark | None cited | 56.6 on tau-Voice, run on Artificial Analysis's harness, without a reasoning step |
| Mechanism | Text generated before audio, enabling live blocking | Text is a "checkpoint": rules read it before speech; tool calls run in parallel with speech |
| Customization | RL-customizable | Fine-tuning via RL and SFT APIs being built with beta customers |
| Funding | "Backed by $35M" | Rime's July 2026 Series A was $24M and its 2025 seed $5.5M, per its own blog; $35M is plausible as total funding but Rime's post does not state it |
Where Prism fits in Rime's own timeline
Rime began as a text-to-speech company founded by linguists. Its blog shows a quick run of releases before this one: Mist v3 in April 2026, Coda in May, a $24M Series A on July 15, a speech-to-text product on October 20 and Coda v2 with phoneme control on October 23. Prism arrives five days after Coda v2, and the Series A announcement already described a move toward speech-to-speech models for regulated enterprises.
Clifford's launch post also contains a notable reversal: eighteen months ago, she writes, she publicly predicted speech-to-speech was a decade from real use. She now predicts that in eighteen months, cascade-based agents "will look like pale, ghastly imitations" of Prism-style agents. That is a forecast from the vendor, and should be read as one.
Three architectures, compared
| Architecture | Strength | Weakness | Example |
|---|---|---|---|
| Cascade (speech-to-text, LLM, text-to-speech) | Fully inspectable and swappable at each stage | Text-only handoff discards tone, hesitation and background sound; stacked latency | Most production voice agents today |
| Fused speech-to-speech | Responsive and expressive | Opaque: no text to inspect before audio; hard to customize | Large-lab realtime models |
| Prism-style text checkpoint | Inspectable text before speech while still hearing the user | New and unproven outside Rime's own tests; cannot reason before speaking | Rime Prism |
The trade-off Rime openly concedes is reasoning: chain-of-thought blows the latency budget of a live call, so none of the fast models think before they talk. Models that do reason score higher on tau-Voice but wait over a second to start speaking.
What Prism does differently
Most large-lab speech-to-speech models fuse text and audio understanding at runtime, which makes them responsive but opaque. Rime's founder and CEO Lily Clifford puts it bluntly: "you can't inspect what it's about to say, you can't use it in a composable system, and you can't own it." The older cascade (speech-to-text, then an LLM, then text-to-speech) is controllable but, she writes, "destroys the most important information" such as hesitation, anger or background sounds.
Prism sits between them. Rime says it is trained end to end as one system doing three jobs: it listens, writes what it will say, and keeps hearing the user while it speaks. The written text is the checkpoint where a business can run guardrails, edit content tags, apply phonemic conditioning for names and account terms, and manage tools.
That design matters for regulated callers. Coverage of the launch notes use cases such as account lookups or freezing a bank card during a live call, with the ability to block a specific reply without pausing the call. If you have followed our reporting on OpenAI's GPT-Live voice or enterprise voice agents from OpenAI Presence, the contrast is that those fuse reasoning and speech, while Prism exposes the text.
The numbers Rime published
Rime ran Prism on the harness Artificial Analysis uses for its speech-to-speech leaderboard, the way it would run live: no reasoning step before speaking. tau-Voice tests tool calling and instruction following on simulated customer-service calls across airline, retail and telecom.
- Overall: Prism scores 56.6 on tau-Voice. Rime says that is 17 to 30 points ahead of every OpenAI and Google realtime model that does not reason first, and ahead of GPT-Realtime-2.1 even with reasoning turned up.
- Telecom: 82.5 versus 45.6 for GPT-Live-1, the domain with the most account and phone numbers to get right.
- Latency: 0.57 seconds to first audio, about half the wait of GPT-Live-1, GPT-Realtime-2.1 and Gemini 3.8 Live.
- Caveat Rime states: two models score above Prism by reasoning before they speak, but start talking after 1.24 to 1.35 seconds. Rime's chart shows GPT-Live-1 (Sol, low) at 59.3% and 1.24s, GPT-Realtime-2.1 High at 45.7% and 1.21s, and Gemini 3.8 Live at 30.1% and 1.18s.
Notably, the chart cites the Artificial Analysis leaderboard as of October 2026 and says Prism was "measured on the same harness", which means Prism's result is Rime's own run, not an independent listing. We could not find a Prism entry on the leaderboard itself. For how that leaderboard has been used, see our coverage of Grok Voice 3.0's speech-to-speech index score and Gemini 3.8 Live with extended thinking.
Why the 300ms number does not hold
The 300ms and "4x faster than GPT-Live" figures come from a news aggregation summary of the launch. Rime's own post says "about 500ms" in the intro and 0.57 seconds to first audio in the benchmark section, roughly 2x faster than the three comparison models, not 4x. We found no Rime page, docs entry or chart that says 300ms. Possible explanations include a different measurement (for example network-free model latency versus time to first audio), or a quote from a social post. Treat 300ms as unverified until Rime or a third party shows how it was measured.
How to check latency claims yourself:
- Measure from end of caller speech to first audible audio on a real telephony path, not model-only time.
- Report the median and the 95th percentile; voice agents fail on the tail.
- Hold reasoning and tool-call settings constant across vendors.
- Test on your own domain; telecom account numbers behave differently from restaurant orders.
The Revmo result, and its limits
Rime shares one customer-style experiment. Using 10 hours of production calls, tool traces and transcripts from Revmo, an ordering-agent company that runs on OpenAI today, Rime built a reinforcement-learning environment and trained a Prism on it. On Revmo calls, 94% of Prism's orders were placed only after the caller said yes to "shall I place it?", versus 57% for OpenAI.
That is a striking gap, but it is Rime's own test on one customer's data, comparing a tuned model against an untuned one. It supports the claim that domain-specific training can change behavior, not that Prism beats OpenAI generally. The same logic is why Rime says the model must be fine-tunable: voice agents cannot reason before speaking without blowing the latency budget, so the intelligence is whatever is in the weights.
What this means if you build voice agents
- If you need guardrails: the text checkpoint is the feature to evaluate. Test whether blocking or rewriting a reply adds noticeable delay on real calls.
- If you run on a large lab's realtime API: compare against your own call transcripts. Rime's numbers are from a simulated benchmark, not your traffic.
- If you care about ownership: fine-tuning APIs are promised, not shipped. Ask Rime for terms before planning around them.
- If you are a safety or compliance team: an inspectable pre-speech layer fits a runtime control model. That matters most for voice agents that can freeze cards or change accounts.
For the broader market, see how Cartesia's Sonic-3.6, Mistral's Voxtral 3 and AssemblyAI's TTS-1 are pushing latency and cost on the text-to-speech side.
A practical evaluation plan before you switch
If you run a voice agent today, a pilot of any new speech-to-speech model should answer a few concrete questions before it touches production traffic.
- Replay real calls. Take a few hundred anonymized calls with known outcomes and compare task completion, not just how natural the voice sounds.
- Stress the hard turns. Account numbers, names, interruptions, silence and angry callers are where benchmarks and demos diverge from the floor.
- Test the checkpoint. Add a rule that blocks a specific phrase and measure the extra delay and what the caller hears when a reply is rewritten.
- Price the whole path. The announcement gives no pricing, so ask for per-minute cost, concurrency limits and data-retention terms, especially for regulated calls.
- Keep a fallback. Route a share of traffic to your current stack so you can compare outcomes side by side.
A microphone linked to a speech bubble, representing a live voice agent call evaluated before a speech-to-speech model goes to production
Open questions
- Where does 300ms come from, and under what measurement?
- Will Artificial Analysis list Prism independently, and does the score hold?
- What does Prism cost, and when do the fine-tuning APIs ship?
- How much latency does a guardrail intervention add in practice?
Related reading
- OpenAI GPT-Live voice in ChatGPT
- Gemini 3.8 Live and Extended Thinking
- Grok Voice 3.0 speech-to-speech
- OpenAI Presence: enterprise voice agents
- Mistral Voxtral 3 and the voice agent SDK
- Cartesia Sonic-3.6 on Artificial Analysis
- Sources: Rime, Introducing Prism, Rime's Series A coverage at TechCrunch
Accurate as of October 28, 2026. Benchmark figures are Rime's own.