Rime Prism vs ElevenLabs: Two Speech-to-Speech Models, One Latency Number

Eleven days apart, two companies shipped the same idea and reported the same number. Rime Prism arrived on October 28, 2026 at 0.57 seconds to first audio. Eleven Realtime v1 Max, ElevenLabs' own end-to-end speech-to-speech model, arrived on October 20, 2026 — also 0.57 seconds. The coincidence is the most useful thing about this matchup, because it means latency is not the decision. Once two models are tied on the clock, what separates them is what happens in the half-second, and these two spend it in opposite ways: Eleven Realtime v1 Max listens and speaks, while Rime Prism writes first.

This is the closest comparison in the batch and the one where the choice actually matters, because both products replace the entire cascade rather than one component, and both price the audio at eight cents a minute.

The short answer

Eleven Realtime v1 Max is the better choice when coverage and voice inventory decide the deal: 90-plus languages and more than 17,000 voices is a catalogue no competitor in this pair can match, and the socket is wire-compatible with OpenAI's Realtime API, which means a migration is an endpoint change rather than a rewrite.

Rime Prism is the better choice when what the agent says has to be inspectable before it is spoken. It is the only model in this pair that hands the application a text string mid-turn with the option to rewrite or block it, and the only one offering on-prem or VPC deployment.

Both are priced at $0.080 per minute of audio, and that apparent parity hides a real gap: ElevenLabs bills $0.080 per minute of connected audio plus model tokens, while Rime bills $0.080 per minute and no tokens. On a call-heavy agent that is not a rounding difference.

What each company actually announced

ElevenLabs shipped a family, not a model. Eleven Realtime v1 Max is the complex-task variant — support, lead qualification, anything with tool use. Eleven Realtime v1 Mini is the conversational variant for assistants, tutors and characters. Both are single end-to-end networks that fold recognition, turn-taking, reasoning and voice generation into one system rather than three stitched stages. ElevenLabs' stated case for doing that is the seam: dead air after a pause, talking over an interruption, a transcript that disagrees with the voice. Those are cascade artefacts, and the fix offered is to tune all four stages together.

The access path is deliberately boring, which is a compliment. Eleven Realtime runs inside ElevenAgents through a dropdown, or through a Realtime API at wss://api.elevenlabs.io/v1/convai/realtime that is wire-compatible with OpenAI's Realtime protocol — change the endpoint and the model name, and an existing OpenAI SDK works unchanged. At launch it is WebSocket only; there is no WebRTC path yet.

Rime's announcement is a different shape. One model, Prism, positioned for enterprise contact centres, with the differentiating claim being the text checkpoint rather than a family of variants. There is a free tier with $50 of credit, Flex at $0.080/min, Scale at $0.064/min, and an Enterprise tier that adds VPC and on-prem plus fine-tuning on customer call recordings and custom voices. No WebRTC either, but streaming, LiveKit and Pipecat integrations, 8 kHz telephony, and SOC 2 with HIPAA claims.

One detail worth flagging for anyone planning a rollout: ElevenLabs describes Eleven Realtime as rolling out gradually, and the model identifiers have existed in its changelog since early August. The October 20 date is the announcement. If your account does not show the model in the ElevenAgents dropdown, that is a staged release rather than a misconfiguration.

The checkpoint is the real difference

Both models are end-to-end. Only one of them leaves a text artifact you can act on before the caller hears anything.

Eleven Realtime v1 Max goes from audio in to audio out. There is a transcript on the other side of the call, and there is a system prompt upstream of it, but there is no point in the turn where your application holds the agent's next sentence as a string and decides whether to send it. That is not a criticism — it is the architecture ElevenLabs chose, and it is the architecture that produces the naturalness the company is selling. Folding four stages into one network is precisely what removes the inspectable seam.

Prism keeps the seam. Rime says the model drafts its reply as text, the application can rewrite, block, tag or fix pronunciation on that text, and the audio follows. Tool calls run in parallel with speech rather than before it. The claim that this is not a cascade rests on the reasoning happening over audio — tone and hesitation and background noise inform the reply — with text as an output waypoint rather than an input.

For a compliance team, that difference is the entire procurement. "The agent cannot say X" is enforceable in Prism at the sentence level and enforceable in Eleven Realtime only at the prompt level, which is to say probabilistically. For a product team building a character or a tutor, the reverse holds: the checkpoint is overhead, and the naturalness that comes from a single tuned network is the product.

Two 0.57-second numbers that are not the same measurement

The parity is real but the instruments differ, and the difference matters if you plan to cite either figure.

Eleven Realtime v1 Max's 0.57 seconds is an Artificial Analysis measurement — a third party ran it, on the same clock used for its other speech-to-speech entries, and published it. Eleven Realtime v1 Mini lands at 0.62 seconds on the same instrument.

Prism's 0.57 seconds is Rime's own figure, produced on the τ-Voice harness that Artificial Analysis also uses, but run by Rime. Prism had no public leaderboard entry at launch. So one number has a referee and the other does not, even though both are the same value to two decimal places. Reporting them side by side as a tie is defensible; reporting them as equivalent evidence is not.

The other scores diverge sharply, and again on different instruments. Eleven Realtime v1 Max posts 98.8% on Big Bench Audio against 75.9 to 76% for Mini, task success of 87.1% against 60.3%, Full-Duplex-Bench of 94.1 against 95.8%, and an Arena Elo around 997 to 998 against 988. It is also unscored on the Speech-to-Speech Index and carries no τ-Voice agentic score — so there is no published number that puts Eleven Realtime v1 Max and Prism on the same axis at all. Anyone who lines up Prism's 56.6 against ElevenLabs' 98.8 is comparing a task-completion benchmark to an audio-quality benchmark and getting a meaningless 42-point gap.

The price gap is in the token line

On the surface both products cost eight cents a minute. Read the rate cards and they do not.

• ElevenLabs — $0.080 per minute of connected audio, plus model tokens. Eleven Realtime v1 Mini is $0.20 input and $6.00 output per million tokens; v1 Max is $1.80 input and $9.00 output. Burst overage doubles the audio rate to $0.16/min. Inside ElevenAgents the standard rate is $0.08/min plus LLM tokens.

• Rime — $0.080 per minute of Prism audio on Flex, $0.064 on Scale, no token charge on top. The free tier carries $50 of usage credit across Prism, Mist v3, Coda v2 and Rime STT.

How much the token line costs you depends entirely on how much the agent thinks. A five-minute support call where the agent takes eight turns and each turn reasons over a few thousand tokens is not free at $1.80 per million input and $9.00 per million output on Max — it is a second meter running alongside the first. For a short acknowledgement-heavy assistant on Mini, the token line is noise. For a Max deployment doing real tool orchestration, it can rival the audio charge.

Rime's flat minute has its own trap, and it is the opposite one. Prism bills call audio rather than agent speech, so a call with a long-winded caller costs the same as one where the agent never stops talking. If your calls skew toward brief agent turns and long customer turns, Prism's per-minute rate is the more expensive of the two by a margin that has nothing to do with the models.

Model both meters on your own call distribution before you commit. Neither vendor publishes a calculator, and the eight-cent headline is genuinely not the number either of them will charge you.

Coverage, voices, and the deployment question

ElevenLabs wins coverage so decisively that it is worth stating plainly rather than burying: 90-plus languages and more than 17,000 voices. Rime does not publish Prism's language count or voice inventory at all, though its separate text-to-speech products publish theirs, and those figures cannot be read across to Prism.

That asymmetry decides a large share of real deployments on its own. A European retailer launching in nine markets does not need to read the rest of this article.

Deployment is where it swings back. Rime offers VPC and on-prem on Enterprise, alongside cloud, plus a co-located endpoint option for its TTS models. Eleven Realtime at launch is a hosted WebSocket service. Rime's argument — that frontier realtime APIs are cloud-only while Prism lets you control where the audio lives — is a real differentiator for health, financial services and any buyer with a data-residency clause in the contract. Rime claims SOC 2 and HIPAA for Prism; ElevenLabs sells enterprise agreements with DPAs and BAAs on its own terms. Both can be made to satisfy a compliance review, and only one of them can be made to satisfy "the audio never leaves our building."

Fine-tuning is the other asymmetry, and it points the same way. Rime says Prism is fine-tunable on a customer's own call recordings, with self-serve RL and SFT APIs in beta. ElevenLabs' tuning surface is voice-level — cloning, professional voice workflows, style and stability settings — rather than a custom acoustic-and-reasoning model trained on your calls. If the goal is "sound like our brand," that is a solved problem on the ElevenLabs side. If the goal is "get better at our specific order flow," only one of these is offering that.

Where OrcaRouter fits, and where it does not

Neither Prism nor Eleven Realtime is on our catalogue. Prism is sold direct by Rime; the Eleven Realtime models are sold direct by ElevenLabs and through several third-party platforms. If you pick either of them you are buying from the vendor, and this article is not going to pretend otherwise.

The reason OrcaRouter still belongs in this decision is that end-to-end speech models do not remove every model from your stack. Both of these products still need a policy layer, a retrieval step, a classifier and an escalation path, and those are ordinary model calls. One key covers 200-plus models at 0% markup over provider list price, so a cheap guardrail model and an expensive escalation model live behind one endpoint with automatic failover — which is the part that matters on a live call, where a stalled provider is indistinguishable from a dropped line. The routing DSL is how you express "keep the acknowledgement local and cheap, escalate the lookup," and that rule is worth writing once in the gateway rather than three times in the agent.

If what you actually want is a cascade with hosted synthesis, openai/tts-1-hd and google/gemini-3.1-flash-tts-preview are on the same key at list price. Different tools, same bill.

How to decide in ten minutes

Three questions, in order, and stop at the first one you answer yes to.

Do you need to operate in more than a few languages, or need a specific voice out of a very large inventory? Eleven Realtime v1 Max or v1 Mini. The catalogue is the product, and Prism does not publish a comparable one.

Does something your agent says have to be approved before a caller hears it, or does the audio have to stay inside your network? Rime Prism. The checkpoint and the on-prem tier are the two things ElevenLabs is not offering here, and they are not features you can add later.

Do neither of those bind, and is the deciding factor cost? Then run both meters on your own call data, because the eight-cent headline is equal and the token line is not. A short-turn, low-reasoning assistant on Mini will beat Prism on price. A tool-heavy Max deployment probably will not.

The thing to watch next is the one number that would settle it. Eleven Realtime v1 Max has no τ-Voice score, and Prism has no independent one. If either model lands on a public leaderboard measured by someone other than its vendor, this comparison gets a great deal shorter. Until then the two are tied at 0.57 seconds on two different stopwatches, and the rest is architecture.

One last line on the numbers above: the Eleven Realtime scores are Artificial Analysis figures, the price cards are the vendors' own, and Prism's 56.6 is Rime's. No instrument in this article puts both models on one axis, which is why the piece is organised around architecture and coverage instead of around a scoreboard.