Two voice-agent APIs arrived around the same time. ElevenLabs published Eleven Realtime on October 20, two speech-to-speech models called v1 Mini and v1 Max that accept the OpenAI Realtime SDK as is. ElevenLabs, Eleven Realtime Mistral’s Voxtral 3 page carries no publication date; its evaluation charts are stamped October 20, so we date the launch to about then. Mistral, Voxtral 3 With OpenAI’s GPT-Live-1 already in the API since September, three vendors now quote a per-minute price for a voice agent, and none of the three prices counts the same things.
S5 Labs has not run any of these models. Every benchmark figure below is the vendor’s, or Artificial Analysis as quoted by the vendor; we did not open the Artificial Analysis pages ourselves.
What each vendor shipped
ElevenLabs’ post, by Min Kim, lists more than 90 languages and more than 17,000 voices from the ElevenLabs library. The code sample uses the Mini ID eleven_realtime_v1_mini; no Max ID appears in the post. ElevenLabs, Eleven Realtime Clients connect to wss://api.elevenlabs.io/v1/convai/realtime, and the post says WebSocket is the only transport. ElevenLabs, Eleven Realtime
Voxtral 3 is a family rather than one model: Small Transcribe 3, Mini Transcribe 3, Mini Transcribe Realtime 3, Mini TTS 3, and a voice agent API that answers and places calls over SIP and runs in browsers and apps over WebRTC. Mistral says each agent turn completes in under one second and describes realtime transcription as sub-500 ms. Mistral, Voxtral 3 The API model IDs are voxtral-mini-transcribe-3-0, voxtral-small-transcribe-3-0, voxtral-mini-transcribe-realtime-3-0, voxtral-mini-tts-3-0 and voxtral-mini-tts-realtime-3-0. Realtime transcription delay is set per session, from 80 to 1,200 ms or 2,400 ms, and includes turn detection. Mistral docs, audio
| Voxtral 3 component | Price (Mistral’s post) |
|---|---|
| Small Transcribe 3 | $0.0040 per minute |
| Mini Transcribe 3 (batch) | $0.0035 per minute |
| Mini Transcribe Realtime 3 | $0.008 per minute |
| Mini TTS 3 | $13 per 1M characters |
| Voice agent API, Mistral Small 4 | $0.04 per minute |
| Voice agent API, larger model (unnamed) | $0.08 per minute |
Source: Mistral, Voxtral 3. The agent prices appear in the post only; we did not find them on a pricing page.
Mistral’s language counts do not agree with each other. The post body says realtime transcription covers “over 35 languages” while the model card figure says 40+. Transcription is listed at 50+, TTS at 35+ languages and 50+ dialects, and the voice presets at 18 languages. Mistral, Voxtral 3
Three per-minute prices that count different things
| Vendor | Headline rate | What it covers | What is extra |
|---|---|---|---|
| OpenAI GPT-Live-1 | $0.05/min, billed per second | The voice layer | Backend model and tool usage, billed separately |
| Mistral voice agent | $0.04/min (Small 4) or $0.08/min (larger model) | The agent, as priced in the post | Nothing stated |
| Eleven Realtime | $0.080 per connected-audio minute | The audio connection | LLM tokens: Mini $0.20 in / $6.00 out per 1M, Max $1.80 in / $9.00 out per 1M |
Sources: OpenAI pricing, OpenAI, gpt-live-1, Mistral, Voxtral 3, ElevenLabs API pricing.
ElevenLabs’ pricing page states the headline as “$0.08+ LLM usage”. We read that as the per-minute charge plus the token charge on top, but the page does not spell out the stacking, so treat it as our inference. The page also lists a burst rate of $0.160 per minute above the concurrency limit, concurrency tiers of 4, 6, 10, 20, 30 and 40, and six rows of included minutes that it calls estimates based on typical token usage: roughly 14 (Mini) or 12 (Max) at the smallest tier, rising to about 11,794 or 9,714 at the largest. The page does not label those rows with plan names, so we are not attaching any. It carries no effective date. ElevenLabs API pricing
Because ElevenLabs publishes no all-in figure, any per-minute total needs an assumption about tokens. Ours: a minute of conversation that sends 1,000 input tokens and generates 300 output tokens on the LLM. That is an S5 Labs assumption, not a vendor number, and the pricing page does not say what kind of token it is pricing.
| Eleven Realtime, modeled | Mini | Max |
|---|---|---|
| Audio, per minute | $0.080 | $0.080 |
| Tokens at 1,000 in / 300 out | $0.002 | $0.0045 |
| Total | about $0.082 | about $0.0845 |
| Tokens at 10,000 in / 3,000 out | $0.020 | $0.045 |
| Total at the higher budget | about $0.100 | about $0.125 |
Arithmetic is ours, from the token prices on ElevenLabs API pricing. The output price dominates on Mini ($6.00 against $0.20 for input), while Max is nine times Mini on input and one and a half times on output, so the gap widens once the agent carries long tool results and histories.
For 1,000 conversation minutes, then: Mistral’s agent on Small 4 is $40 and on the larger model $80; GPT-Live-1 is $50 before any backend model or tool call; Eleven Realtime is $80 before tokens and $82 to $125 under our two token budgets. Our GPT-Live-1 coverage walks through what that separate backend line can include. For teams on OpenAI’s Realtime API rather than GPT-Live, the pricing page lists only tokens, with no per-minute rate: gpt-realtime-2.1 audio at $32 in / $64 out per 1M (text $4 / $24, cached $0.40), gpt-realtime-2.1-mini audio at $10 / $20, and gpt-realtime-2 audio at $32 / $64. OpenAI pricing
Migrating off OpenAI Realtime
ElevenLabs’ pitch is that the OpenAI SDK works unchanged because the API follows the OpenAI Realtime specs. ElevenLabs, Eleven Realtime The constraint is transport: the API is WebSocket only, so a WebRTC integration has to move to WebSocket before it can point at api.elevenlabs.io. ElevenLabs, Eleven Realtime We have not confirmed event-level compatibility from the ElevenLabs docs, so treat “works as is” as the vendor’s claim until your own client has run against it. We also did not recheck whether the product runs separately from ElevenAgents.
OpenAI’s own newest voice model is not a Realtime drop-in either. The gpt-live-1 model page lists v1/live/sessions as its only supported endpoint, with Chat Completions, Responses and Realtime unsupported; function calling is supported. OpenAI, gpt-live-1 A team that built on Realtime and wants GPT-Live-1 is changing endpoint and architecture anyway, which is the gap ElevenLabs is aiming at.
Mistral’s agent speaks SIP and WebRTC; the WebSocket client in its SDK (the realtime extra) is for realtime transcription, not the agent. Mistral docs, audio Mistral, Voxtral 3 We found no OpenAI-compatibility claim in Mistral’s material, but we did not check it exhaustively, so count on a rewrite rather than a URL swap.
Max or Mini for tool calling
ElevenLabs reports five comparisons between its two models. Three were run by Artificial Analysis on public benchmarks; the other two carry no named evaluator and should be read as ElevenLabs’ own figures.
| Measure (per ElevenLabs’ post) | Max | Mini | Who ran it |
|---|---|---|---|
| Big Bench Audio (1,000 questions) | 98.8% | 75.9% | Artificial Analysis |
| Full-Duplex-Bench | 94.1% | 95.8% | Artificial Analysis |
| Speech Agent Arena | 998 | 988 | Artificial Analysis |
| Time to first audio | 0.57 s | 0.62 s | Not named |
| Tool-use task completion | 87.1% | 60.3% | Not named |
Source: ElevenLabs, Eleven Realtime.
Mini wins Full-Duplex-Bench by 1.7 points, and the two sit ten points apart on Speech Agent Arena. The nearly 27-point tool-use gap is the number that would decide a purchase, and it is the one with no named source. If an agent’s job is to call tools, reproduce that gap in your own harness, with your own schemas, before paying Max’s token rates.
Mistral’s hosting and the open-weights question
Mistral says Voxtral 3 is “available today through Mistral’s EU-hosted API or for self-deployment on your own infrastructure.” The same page lists US and global endpoints, so EU hosting is an option, not the only one. On-prem or private cloud means contacting sales. Mistral, Voxtral 3
“Self-deployment” does not establish open weights, and we could not confirm any for Voxtral 3. Neither the post nor the docs say anything about weights or a license, and the Hugging Face Voxtral collection holds five items, none of them Voxtral 3: Voxtral-4B-TTS-2603, Voxtral-Mini-4B-Realtime-2602, Voxtral-Mini-4B-Realtime-Arabic, Voxtral-Small-24B-2507 and Voxtral-Mini-3B-2507. Hugging Face, Voxtral collection The prior generation is what you can download today. Voxtral-Mini-4B-Realtime-2602 is Apache 2.0, covers 13 languages and recommends a 480 ms delay; its card disagrees with itself on the lower bound of the configurable range, 80 ms in one place and 240 ms in another. Model card Voxtral-4B-TTS-2603 is CC BY-NC 4.0 with nine languages, so it is not a commercial option as licensed. Model card Our open-source models guide tracks the license terms on the earlier Voxtral checkpoints alongside other open-weight models. The Arabic realtime checkpoint is also the natural comparison for TII’s Falcon ASR if Arabic is in scope.
On accuracy, Mistral’s page charts FLEURS across 45 languages, realtime FLEURS across 35, and diarization cpWER of 18.4% for Small and 18.6% for Mini. An Artificial Analysis WER Index chart appears without a score we could read, and the TTS judge is printed as “gemini-3.8-flash”. Mistral, Voxtral 3
October 28 update: a self-hosted voice stack from Ollama and llama.cpp
Three releases in five days make a local voice stack buildable. Ollama v0.40.11 (October 24, UTC) generates speech on macOS with the phonon model through the OpenAI-compatible /v1/audio/speech endpoint; the demo is ollama run phonon "Hello from Ollama.". Ollama v0.40.12 (October 27) added lfm3-audio for speech recognition, translation and text to speech in 18 languages. llama.cpp v0.8.0 (October 28) gave llama-server a /v1/realtime WebSocket on the OpenAI realtime protocol, plus POST /tts, and lists LFM3-Audio-2B and LFM3-Omni-2B for ASR, TTS and voice chat from one model. We have not run any of it. Ollama v0.40.11 · Ollama v0.40.12 · llama.cpp v0.8.0
Ollama’s notes say macOS, and the phonon library page requires Apple Silicon and Ollama 0.40.9 or later. The speech docs list Linux and Windows playback commands and describe writing a WAV on servers without speakers, but never say the runtime works there, so we read Ollama’s phonon as Apple Silicon only. The model is not: Gradium ships its own runtime and claims the 31M checkpoint runs on a laptop CPU or in a browser tab. The v0.40.12 note limits lfm3 and d1 2b to Apple hardware and does not name lfm3-audio. Ollama v0.40.13, published later on October 28, says the LFM3 models, including lfm3-audio and lfm3-omni, now run on Linux and Windows as well as Apple silicon; its commits also route /v1/audio/speech on GGUF models to llama-server’s /tts and bump llama.cpp to b11963. Its notes say nothing about phonon on those systems. Ollama, phonon · Ollama docs, speech · Gradium model card · Ollama v0.40.13 · Ollama PR #19344
The endpoints differ. Ollama’s returns WAV (16-bit mono, 24 kHz) or raw PCM, takes up to 4,096 characters per request, and is text to speech only, with no cloning. Its library page and docs disagree on which voices the default tag carries, so run ollama show phonon before promising one. llama.cpp’s /tts returns raw audio bytes and takes a speaker reference for cloning, which the docs say to always provide; it is not the OpenAI shape. Its /v1/realtime transcribes with streaming-encoder models such as Voxtral Realtime and holds tool-calling conversations with LFM3-Audio and LFM3-Omni, the client running each call. The README describes no server-side turn detection, no barge-in beyond a client response.cancel, and no SIP or WebRTC. Ollama docs, speech · llama-server README · llama.cpp PR #31287
| Component | License | Commercial use |
|---|---|---|
| phonon 31M English (Ollama’s default tag) | Apache 2.0 | Card says yes, voices in that repo included |
| phonon 40M (en, fr, de, es, pt, multilingual) | CC BY-NC 4.0 | No |
| LFM3-Audio-2B, LFM3-Omni-2B | LFM Open License v1.0 | Only below $10M annual revenue, counted across the company and its affiliates |
| llama.cpp | MIT | Yes |
| gradium/phonon-voices catalog | Not checked | Unknown |
Sources: Gradium model card, Ollama, phonon, LFM3-Audio-2B license, llama.cpp license. Our LFM3 2B family article has more on the Liquid checkpoints.
Against the hosted rates above, the self-hosted line is $0 per minute plus hardware, power and someone to run it; that is our arithmetic, not a vendor figure. The only hardware numbers in the sources are phonon’s loaded memory on an Apple M5 Max: 121 MB (31M) and 141 MB (40M). What you give up: phonon speaks but does not listen or think, so the LLM is a separate process; LFM3-Audio and Omni answer and call tools from one 2B model, but the card text on Ollama’s page advises against agentic coding or complex multi-step reasoning; nothing here answers a phone call; and no source gives throughput, concurrency or a quality comparison with the hosted agents. Ollama, lfm3-audio · Ollama, phonon
Two footnotes. The v0.8.0 notes list “--model or -hf is now required” under API changes, but the merged change only reworded an error message; a launch without a model already failed, so only scripts that match on the old text need touching. And Rime launched Prism, a speech-to-speech model it says anyone can fine-tune, on October 28; the launch page publishes no price. llama.cpp PR #31242 · Rime, Prism If you serve speech models through Xinference instead, read our notes on its 4.1.0 security fixes first: a CosyVoice config file can run code on the worker.
What is missing before you can price this
- ElevenLabs: no effective date on the pricing page, no published all-in cost per minute, no Max model ID in the post, and no confirmation of event-level Realtime compatibility.
- Mistral: no date on the Voxtral 3 post, agent prices that appear only there, an unnamed larger model behind the $0.08 rate, and no stated license or weights.
- OpenAI: the gpt-live-1 model page carries no release date, and the backend cost that sits beside the $0.05 line depends on which model does the reasoning.
- Across all three: no independent run on one shared harness. Artificial Analysis covered three Eleven benchmarks; nothing we found compares the three agents on the same calls.
For a cost-sensitive agent that does not lean on tools, Mistral’s $0.04 line on Small 4 is the lowest published number and the only one that names no separate charge; confirm what it includes before building a budget on a blog post. For a team already on OpenAI Realtime, Eleven Realtime is the smallest code change on offer, with the token line as the variable to measure. A hundred recorded calls, with tokens per minute and tool success logged per model, would replace our token assumption with your own number.
Sources
Vendor posts and docs: ElevenLabs, Eleven Realtime · ElevenLabs API pricing · Mistral, Voxtral 3 · Mistral docs, audio · OpenAI pricing · OpenAI, gpt-live-1 · Ollama docs, speech · Ollama, phonon · Ollama, lfm3-audio · llama-server README · Rime, Prism
Releases and pull requests: Ollama v0.40.11 · Ollama v0.40.12 · Ollama v0.40.13 · Ollama PR #19344 · llama.cpp v0.8.0 · llama.cpp PR #31242 · llama.cpp PR #31287
Model cards and licenses: Hugging Face, Voxtral collection · Voxtral-Mini-4B-Realtime-2602 · Voxtral-4B-TTS-2603 · Gradium, phonon-31m-en · LFM3-Audio-2B license · llama.cpp license