
MAI Realtime 1 vs Eleven Realtime Max: An Eightfold Price Gap, and the One Number That Decides It
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 121 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 119 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 143 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 920 tok/s
The published rates make this look settled. MAI Realtime 1 costs $0.01 per audio minute. Eleven Realtime Max — ElevenLabs' top speech-to-speech model, listed as Realtime v1 Max — costs roughly $4.80 per hour on Artificial Analysis' independent normalisation, which is $0.08 a minute. Eight to one, before you count anything else. That comparison is also misleading in a specific, arithmetic way, and the arithmetic is the whole article: MAI Realtime 1 bills a second, token-priced reasoning model on top of its minute rate, and nobody outside Microsoft has published how often that second meter runs.
So the useful question is not which headline is smaller. It is which of these two you can cost before you sign, and which one you have to measure first. Those answers point in opposite directions, which is why this is a real decision rather than a formality.

What each one is, in one line
• MAI Realtime 1 — Microsoft's first speech-to-speech model, announced 17 November 2026 at Ignite. Two tiers: a fast conversational path at a cent a minute, and a reasoning engine at $1.00 per million input and $5.00 per million output tokens that harder turns hand off to. Limited-access preview via Foundry; demo in MAI Labs. Speaks the OpenAI Realtime API.
• Eleven Realtime Max — ElevenLabs' speech-to-speech model for voice agents doing harder work than small talk. One tuned system covering recognition, turn-taking, an LLM and voice. Available in ElevenAgents and through an OpenAI-compatible Realtime API, WebSocket only. Covers 90+ languages.
• What neither publishes — languages supported by MAI Realtime 1. Microsoft's launch material lists none, and for a contact-centre product that is a first-order question, not a footnote.
The one side with real measurements
Eleven Realtime Max is well covered by independent testing, and the profile is lopsided in a way that matters for agent design:
• Speech-to-speech index — 77.3, from Artificial Analysis, on a composite of reasoning, conversational dynamics and agentic ability.
• Speech reasoning (Big Bench Audio) — 98.8%. That is among the highest figures on the board, and it is the number the model is built around: it answers hard spoken questions accurately.
• Time to first audio — 570 milliseconds, the fastest of the two on a straight comparison and fast enough to feel immediate.
• Agentic task completion (τ-Voice) — 47.6%. This is the soft spot. A model that reasons beautifully about what it hears still finishes fewer than half of the multi-step tasks the harness sets it.
• Conversational dynamics — 94.1%.
• Blended cost — about $4.80 per hour of audio, per Artificial Analysis' normalisation of a 40-question Big Bench Audio subset to an hourly basis. That is a measurement convention, not ElevenLabs' rate card, and the two should not be quoted interchangeably.
MAI Realtime 1 has almost none of these, and the gap between "almost" and "none" is the interesting part. Microsoft publishes one accuracy number on the model page: 98.8% on Big Bench Audio. It is a vendor figure — Microsoft's own evaluation, on a benchmark Microsoft picked, with no methodology published beside it — and no independent board has re-run it. Note also that 98.8% is arithmetically identical to the Eleven Realtime Max figure above, which is a coincidence of two different harnesses landing on the same percentage and should not be read as the two models tying. Beyond that single number, the page carries a P50 time-to-first-audio card and a "P50 Latency on BigBench Audio" caption, and nothing else. No τ-Voice score, no language list, no distribution behind the percentile. Microsoft's larger claim — that the model sits on the performance-price Pareto frontier for Big Bench Audio and τ-Voice — is a claim about where a point would sit on a chart that has not been drawn in public. Artificial Analysis does not carry the model at all as of 18 November.
That asymmetry is not an argument for ElevenLabs. It is an argument about what you can know today: one of these models has a measured 47.6% agentic score you can plan against, and the other has a vendor-reported accuracy figure and a percentile.
A note on sourcing before the numbers start moving: every Eleven Realtime Max performance figure below is Artificial Analysis' independent measurement. The one MAI Realtime 1 performance figure quoted anywhere in this piece — 98.8% on Big Bench Audio — is Microsoft's own, unaudited, and is labelled as such wherever it appears. Nothing else about its performance is published at all.
The arithmetic nobody has finished
Start with the voice line, because it is the one thing both sides publish honestly.
• MAI Realtime 1 — $0.01/min → $0.60 per hour of conversation, fast path only.
• Eleven Realtime Max — about $0.08/min equivalent → $4.80 per hour, on Artificial Analysis' basis.
An eightfold gap on the minute meter. Now add the second meter, and note what is missing from the public information: Microsoft has not said what fraction of turns escalate to the reasoning tier. Every cost model you build therefore has an input you have to invent. Here is what the invented input does to the answer. Take a five-minute support call — $0.05 of voice time — and assume that across its escalated turns the reasoning engine sees 5,000 input and 1,500 output tokens in total. At Microsoft's $1.00 and $5.00 per million rates that is $0.0125 of reasoning for one full pass over the call.
• Zero escalation — $0.050 per call against ElevenLabs' $0.40. Eight to one, as advertised.
• One full reasoning pass — $0.063 per call. Still under a sixth of the ElevenLabs figure.
• Three full reasoning passes — $0.088 per call. Just over a fifth.
• Ten full reasoning passes — $0.175 per call. About 44% of the ElevenLabs figure, and this is a workload no support conversation plausibly generates.
Read that again: even ten times the reasoning load does not close an eightfold gap on the voice line. Which means the honest version of this comparison is that MAI Realtime 1 is cheaper by a wide margin under every escalation assumption a real conversation supports — and the reason to hesitate is not price. It is that you cannot yet tell whether the model is good enough for the job, and Microsoft's preview is unbilled until public preview, so finding out costs nothing but engineering time.
Two caveats on that arithmetic, both of which cut against treating it as a forecast. The reasoning estimate is mine, not Microsoft's — token counts vary enormously between a lookup that needs one tool call and an exception that needs four. And the ElevenLabs side uses Artificial Analysis' hourly normalisation rather than ElevenLabs' own rate card, because ElevenLabs does not publish a flat per-minute dollar figure for Realtime Max; it prices minutes plus LLM tokens with an in-app estimate. So the comparison is directionally solid and numerically soft on both sides, which is normal for voice and worth saying out loud.
One point does deserve its own line. The Foundry limited-access preview does not bill customers until public preview begins. A cost comparison written in November is therefore a comparison of what you would pay later, against a preview you can run now at zero. The rational move is to measure, not to negotiate.
Where ElevenLabs is not losing on price at all
Two things the eightfold figure hides.
The first is that ElevenLabs is not selling a minute meter. ElevenAgents prices per minute plus LLM token costs, with an in-app estimate per minute, and the Realtime API charges per minute of connected audio plus tokens at the same token prices on every plan. A team already inside ElevenAgents is not comparing $0.01 to $0.08 — they are comparing a marginal change against a platform they have already integrated, with phone, web widget, WhatsApp and contact-centre deployment, plus SOC 2, HIPAA, GDPR, EU data residency and zero-retention options.
The second is that the compliance list is a product feature with a price of its own. For a clinical or financial voice agent, EU data residency and zero retention are not nice-to-haves that a cheaper model can substitute for. Microsoft's announcement says nothing about data residency, retention or certification for MAI Realtime 1 — and the model is in limited preview, which is exactly the stage at which those terms are usually unset.
So the split is not "cheap versus expensive." It is "cheap and unmeasured, with enterprise terms still unwritten" against "four times the price with the paperwork done and a 47.6% agentic score you can decide is or is not enough."

Running both without choosing yet
Both models speak the OpenAI Realtime API, which is the single most useful fact in this comparison. A voice stack written against that interface can point at either one, which turns the decision into a configuration rather than a rewrite. That is also the reason to keep both reachable during evaluation rather than picking first and testing second.
This is the case OrcaRouter is built for, and it is worth being precise about which half applies. MAI Realtime 1 is not in our catalogue — it is reachable through Microsoft's MAI Labs demo and the Foundry limited-access preview, full stop. What we route is the layer behind it: the reasoning models a voice agent escalates into, at provider list price with 0% markup, so a rate change on that tier reaches your invoice the day it happens rather than at renewal. Automatic failover matters more in voice than anywhere else, because a stalled turn is a human listening to silence; a request that moves to a second provider in a few hundred milliseconds is inaudible, and one that retries for four seconds is the thing customers complain about. When the preview ends and MAI Realtime 1 becomes generally available, the same key and the same routing rules carry it — provided it lands in the catalogue, which today it has not.

Who should do what this week
If your voice agent is a support or scheduling surface where most turns are short and the hard ones are lookups, MAI Realtime 1's economics are not close, and the preview is free — build the evaluation now, instrument the escalation rate yourself, and treat Microsoft's unpublished tail latency as the thing your own measurement replaces.
If your agent's value is in the multi-step work — booking against three systems, verifying identity, resolving an exception — then the 47.6% agentic score on Eleven Realtime Max is a benchmark you can read and MAI Realtime 1's is a blank. Buying on price alone here means buying an unmeasured model for the hardest part of the job.
And if compliance decides your vendor before performance does, nothing in Microsoft's announcement changes your answer this month. Limited preview is not a place to run protected data, and the terms that would let you are not published.
The eightfold gap is real and probably durable. It is also the least interesting number here, because the question that decides this matchup is whether a cent-a-minute model can finish a task — and the only party who has not told us is the one selling it.
