Rime Prism Writes Before It Speaks. That Is the Whole Product.

Rime Prism, launched on October 28, 2026, is a speech-to-speech model for enterprise voice agents, and the feature it will be judged on is not voice quality. It is that Prism emits its reply as text before it says it aloud, so your application can read, rewrite or block the sentence while the call keeps running. Rime bills this as the first fine-tunable frontier speech-to-speech model and prices it at $0.080 per minute of call audio on the Flex tier, $0.064 on Scale. The vendor's headline benchmark number is 56.6 on τ-Voice at 0.57 seconds to first audio — a figure Rime ran on the harness Artificial Analysis uses, not a published leaderboard entry — and the model that outscores it in Rime's own chart, GPT-Live-1, needs 0.67 seconds longer before it starts talking. Both halves of that sentence matter, and the second half is the one nobody put in their headline.

What the text checkpoint actually buys you

Every voice agent built on a cascade — speech recognition, then a language model, then text-to-speech — already has text in the middle. Every voice agent built on a fused realtime model from a big lab does not: it hears audio, thinks, and produces audio, and the only inspectable artifact is a transcript that arrives after you have already spoken the words. Rime's argument is that the first group has control and the second group has audio fidelity, and Prism is an attempt to keep both.

In practice the checkpoint is a string that surfaces mid-turn. Rime says the text layer supports four things an application can act on before audio goes out:

• Rewriting — replace a sentence the policy layer rejects, and the model speaks the replacement rather than the original.

• Blocking — a guardrail that finds a prohibited claim, a competitor name, or an unapproved disclosure can stop that utterance outright while the call continues.

• Tagging — action and content tags ride along with the text, so an orchestrator can trigger a tool call or a hand-raise without waiting for the audio to finish.

• Pronunciation control — names, account numbers and product terms get fixed in text, which is where "spell it as a code, not a word" actually belongs.

And tool calls run in parallel with speech rather than in the gap before it, which is the structural difference from a cascade: the lookup does not occupy the silence between the caller finishing and the agent starting.

Rime is explicit that this is not a cascade in disguise. It insists the reasoning happens over the audio itself — tone, hesitation, laughter, background noise — and that the text is an output waypoint rather than the model's input. You do not have to take the claim at face value to see why enterprise buyers care: the checkpoint is the artifact an audit trail needs. "What did the agent say, and who approved it" becomes a query against a retained string and a policy decision, not a reconstruction from a call recording.

The benchmark claim has a hole in it, and it is easy to check

Here is the chart Rime published, as Rime published it:

• Rime Prism — 56.6 τ-Voice, 0.57 s time to first audio, no reasoning step before speech.

• GPT-Live-1 (Sol, low) — 59.3 τ-Voice, 1.24 s.

• GPT-Realtime-2.1 High — 45.7 τ-Voice, 1.21 s.

• Gemini 3.8 Live — 30.1 τ-Voice, 1.18 s.

Look at the top line. Rime's own chart shows a model scoring higher than Prism. The launch framing — "ahead of non-reasoning realtime models from OpenAI and Google by 17 to 30 points" — is true as scoped, and it is scoped around precisely the model that beats it. GPT-Live-1 lands 2.7 points above Prism and pays 0.67 seconds more to start talking. That is a legitimate Pareto trade and Rime is not hiding it; the number is right there on the page. But the sentence travels further than the chart does, and the sentence is what gets quoted.

There is a second internal tension. Rime also reports 82.5 on the telecom subset of τ-Voice against 45.6 for GPT-Live-1. A 37-point gap on one domain and a 2.7-point deficit on the overall run are not impossible — domain weighting and task mix can do that — but they cannot both be read as "Prism is better at telecom and worse at everything else," and Rime does not break out the subset sizes. Until it does, treat the telecom number as a marketing subset, not a capability claim.

There is a third figure floating around that is not a benchmark at all. Rime's launch material talks about roughly 300 ms of "responsiveness." The measured number in the chart is 0.57 seconds to first audio. Those are different measurements — one is internal responsiveness, one is the time a caller actually hears a voice — and quoting the smaller one next to GPT-Live-1's 1.24 s would be comparing a stopwatch to a lap time.

None of this is a verdict on Prism's quality. It is a note on what the evidence currently supports. Rime ran Prism on the same harness Artificial Analysis uses and had not, at launch, appeared on the public leaderboard. The τ-Voice benchmark itself is real and independent — it comes out of a Paper under review at ICML 2026, built on 278 airline, retail and telecom service tasks with a voice user simulator that varies accent, background noise and turn-taking — and its authors found voice agents landing well below text agents on the same tasks. That context is what makes 56.6 interesting rather than disappointing. It is also what makes independent replication the thing to wait for.

What Prism costs, and what a cascade costs

Prism is billed per minute of call audio, which is a billing unit most TTS vendors do not use. The published tiers:

• Free — $50 of monthly usage credit, no card, and Prism is included in it.

• Flex — $0.080 per minute of Prism audio, no monthly fee, no commitment. Rime STT $0.002/min, Mist v3 $0.030/min, Coda v2 $0.050/min.

• Scale — $100/month with $100 of credit rolling over to $200, and a 20% discount that puts Prism at $0.064/min.

• Enterprise — custom, annual, and the only tier that adds VPC and on-prem deployment plus fine-tuning on customer calls and custom voices.

At Flex rates, a thousand minutes of Prism audio is $80 in model cost before telephony, workflow and platform fees. Whether that is expensive depends entirely on what it replaces. A cascade pays three bills: recognition, generation, synthesis. A typical mid-market stack might spend two to three cents a minute on speech recognition, a few cents on the language model depending on how much it thinks, and three to five cents on synthesis — so the cascade total often lands near or above Prism's eight cents, and that is before you count the latency of three hops or the engineering time to make them behave like one system.

The comparison flips if your language model of choice is a cheap one and your synthesis vendor is cheap. At $0.030/min for Mist v3 on the same Rime account, a lean cascade can undercut Prism substantially — the price of that saving is losing the checkpoint and inheriting the seams. Which is the honest shape of the decision: Prism is not selling intelligence at a premium, it is selling a reviewable text layer at a premium, and you should only pay it if you need the review.

Who this is built for, and who should wait

The design partners named at launch — Dialpad as a design partner, with Asurion, Upstart and Mayo Clinic named as customers and Twilio Ventures among the investors — point at the buyer Rime is chasing: contact-center platforms and telephony vendors who already have compliance obligations attached to every word their agents say. Rime claims SOC 2 and HIPAA, calls Prism telephony-ready, serves it over a WebSocket, and offers cloud, VPC or on-prem deployment. That last one is the quiet differentiator, because the frontier realtime APIs are cloud-only.

The customer evidence is thinner than the feature list. Revmo, an ordering agent, fine-tuned Prism on about ten hours of production calls and reported that 94% of orders were placed only after the caller agreed to "shall I place it?" — against 57% for the OpenAI agent it replaced. Rime also cites a national restaurant chain at 8% more completed orders, 4x fewer hangups across customers, and a Fortune 500 financial services company at 42% higher containment. None of these disclose sample size, the specific OpenAI model, or the measurement method. They are the kind of numbers that are directionally interesting and evidentially weak, and the fine-tuning on ten hours of domain audio is doing most of the work in the Revmo case — a base Prism without that tuning is not what was measured.

Wait if you need multilingual coverage today. Rime does not publish Prism's language list or voice count, while its separate text-to-speech products do publish theirs. Prism-specific numbers are absent from the product page, which is not the same as bad coverage, but you cannot plan a 30-market rollout on an unpublished list.

Wait too if you need a benchmark you can cite internally. There is no independent Prism score yet. If your procurement process requires a third-party evaluation, the honest position this week is "vendor-run on an independent harness, awaiting publication," and you should say it in exactly those words.

The checkpoint is a model call, and that is where routing comes in

One thing the launch material underplays: the text checkpoint is only useful if something can act on it fast enough not to break the conversation. A policy classifier, a claim-detector, a brand-safety pass — each of those is a model invocation sitting between the draft and the audio. That is a latency budget inside a latency budget, and it is the part of the design that decides whether Prism feels instant or feels like a pause.

Prism is sold direct by Rime and is not on OrcaRouter's catalogue. What is on our catalogue is the other half of the problem: the models a checkpoint calls. One key covers 200-plus models at 0% markup over provider list price, so a small fast classifier and a larger escalation model sit behind the same endpoint, with automatic failover if one of them stalls mid-call — which is exactly the failure mode that turns a clever guardrail into dead air. If you are running Prism in front of an existing cascade instead, the synthesis side is here too: openai/tts-1-hd and google/gemini-3.1-flash-tts-preview are both routable today.

The routing DSL matters more than it sounds in this context. When the checkpoint's job is "send this sentence to a cheap model unless it contains a number above $500 or a regulated term," that is a routing rule, and writing it in the gateway beats writing it in the application's hot path.

What to watch, and what would change the answer

Three things would move Prism from an interesting launch to a defensible default, and each is checkable.

The first is an Artificial Analysis entry. Vendor-run on the right harness is a reasonable start; a public leaderboard row is what turns 56.6 into a number you can put in a slide without an asterisk.

The second is the fine-tuning APIs leaving beta. Rime says self-serve RL and SFT APIs are in beta. The Revmo result — the strongest evidence in the launch — came from ten hours of domain audio. If that gap is real and reproducible, it is the actual product. If it is a single customer with an unusually clean dataset, Prism is a good checkpointed model and not a category change.

The third is the language list. A speech-to-speech model for enterprise contact centers that covers eight languages is a regional product; one that covers forty is a different company's procurement decision entirely.

Until those land, the fair summary is this: Rime shipped a genuinely different control surface for voice agents, priced it competitively against the cascade it replaces, and wrapped it in a benchmark chart that is more honest than its press framing. The checkpoint is real. The leaderboard position is a claim. If your agents say things that have to be approved before a customer hears them, the first fact is enough to justify a trial. If your agents mostly need to sound good in twenty languages, it is not.

The one-line version for people who skim

Prism is a $0.08-per-minute speech-to-speech model that hands you the sentence before it speaks it, at a latency cost of roughly half a second from the moment the caller stops talking. It is not cheaper than a lean cascade and it is not top of any public leaderboard. It is the only thing shipping today that gives a compliance team a place to stand between the model's decision and the customer's ear — and that is a feature you either need badly or not at all.

If you want to price the alternatives before you commit, the cascade side of the catalogue is a single endpoint away, and the per-minute arithmetic above is the number to bring to the trial.