
Ming-3.0-audio-flash vs Eleven Realtime v1 Max: Two 99s, Only One of Them Costs $0.08 a Minute
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 185 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 228 tok/s
- 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 · 98 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 79 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 118 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 · 94 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 · 199 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 113 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Ming-3.0-audio-flash and Eleven Realtime v1 Max both claim a speech-reasoning score of about 99%, and the similarity ends the moment you look at how each number was produced. ElevenLabs puts "99% on speech reasoning, 570ms for time to first audio" on its own pricing card for Realtime v1 Max; Artificial Analysis measures the same model at 98.8% and 0.567 seconds. inclusionAI's card claims 99.40 on BigBenchAudio for Ming-3.0-audio-flash; Artificial Analysis measures 99.8% on Big Bench Audio and 1.12 seconds. One vendor undersold its own model by a rounding step and the other undersold its by four tenths. Both are close enough to be credible and neither is the number you should make a decision on.
The decision number is different for each. For Eleven Realtime v1 Max it is $0.08 a minute plus LLM usage, which is roughly $4.80 an hour of input audio at Artificial Analysis's measured rate. For Ming-3.0-audio-flash it is four 80 GB H800s, because there is no rate card at all — it is a 126-billion-parameter mixture-of-experts checkpoint under MIT that you download and serve. Everything else in this comparison follows from that split: one is a subscription to a voice agent, the other is a machine you own.
What each one actually is
Eleven Realtime v1 Max is not a model in the checkpoint sense. The documentation describes each Realtime model as a complete voice agent — speech recognition, voice activity detection and turn-taking, an LLM, and text to speech — packaged so you select it rather than assemble it. Max is positioned for precise, thoughtful responses across multi-step tasks and is the only one of the two Realtime models that lists tool calling. It speaks 90+ languages, is reachable through an OpenAI-compatible Realtime API, and switching between Max and its smaller sibling is a model selection change rather than a rebuild.
Ming-3.0-audio-flash is the opposite arrangement. An audio encoder and projector feed a mixture-of-experts language backbone; inputs are audio plus a text instruction; outputs are text, or text with synthesised speech through an optional Talker. There is no built-in turn detection, no session management, no tool-calling layer described in the card. What the card does claim is breadth on the input side: Chinese and English including accents and dialects, non-speech sound, acoustic scenes, music, and long recordings.
• Parameters — Ming-3.0-audio-flash: 126B total, 6.5B active per step. Eleven Realtime v1 Max: not disclosed.
• Shape — Ming: a checkpoint you serve. Eleven: a hosted agent you select.
• Context — Ming: 131,072 tokens declared in the config. Eleven: not published; the session is the unit, not the token window.
• Languages — Ming: Chinese and English. Eleven: 90+ per the Eleven v4 language list.
• Voice output — Ming: optional, via the Talker, with built-in and reference voices. Eleven: the product's core; voice is chosen from a library.
• Tool calling — Ming: not described. Eleven: listed for Max, not for Mini.
• Licence — Ming: MIT on the weights. Eleven: a commercial API agreement.

The two numbers, and why they disagree in opposite directions
The ElevenLabs card's "99%" is almost certainly a rounded 98.8 from an Artificial Analysis run, published on the vendor's own pricing page with the latency figure attached. That is unusually forthcoming: most vendors quote a benchmark the lab that built the model chose. It also means the two sources are not independent — the vendor is citing the third party, not running a parallel harness — so when they agree, they agree by construction.
The Ming figure is a different case. The October model card carried an evaluation image showing 99.40 on "BigBenchAudio" alongside a much weaker long-audio picture. Today's 99.8 comes from Artificial Analysis directly, and the card was edited this morning to quote it. The gap between 99.40 and 99.8 is not a correction; it is two runs of a benchmark with the same name and different harnesses, prompts and item subsets. Neither is wrong. Both are the kind of number that moves by tenths when the details change, which is why a one-point lead at the top of a table where 42 models cluster in the nineties is not a moat.
What the leaderboard ordering does establish is membership in the top group. On speech reasoning, Ming-3.0-audio-flash at 99.8 and Eleven Realtime v1 Max at 98.8 are separated by one point and by nothing else the board measures. On time to first audio they are separated by a factor of two: 1.12 seconds against 0.567. For a model that has to answer inside a conversational pause, that is the number that decides whether the system feels alive, and it is the one place where the cheap-looking option is not obviously the worse one.
Where the comparison breaks, and it breaks on conversation
Speech reasoning is multiple-choice questions delivered as audio. A voice agent is a different problem: knowing when to start talking, when to stay quiet through a pause, how to take an interruption, whether to emit a backchannel. Artificial Analysis measures that separately on Full Duplex Bench, and this is the least symmetric part of the whole matchup.
Eleven Realtime v1 Max has a published Full Duplex Bench score of 0.941, with interruption handling at 0.99, turn-taking at 1.0, pause handling at 0.978 and backchannels at 0.796. It also has a tau-voice score of 0.578 across airline, retail and telecom task simulations — a much less flattering number, and a useful corrective to anyone reading the 99% as a general verdict. Ming-3.0-audio-flash has no Full Duplex Bench score, no tau-voice score, and no published measurement of how it behaves when someone talks over it. It is not that the checkpoint would fail those tests. It is that nobody has run them, and the model card does not describe the machinery that would make them meaningful: turn detection and barge-in are things you build around it, not things it ships with.
That asymmetry has a practical shape. Choosing Eleven Realtime v1 Max buys you a solved conversational loop with a published (if middling) task-completion score. Choosing Ming-3.0-audio-flash buys you an unusually strong ear and a text interface, and leaves the loop — VAD, endpointing, interruption handling, the LLM turn behind the voice — as your work. The two products are not substitutes at the same layer, and a head-to-head on accuracy quietly implies they are.
The arithmetic of an hour, done honestly
An hour of continuous input audio on Eleven Realtime v1 Max bills at $0.08 a minute, which is $4.80, plus whatever the bundled LLM consumes. Artificial Analysis's measured cost per hour of input audio for the model is exactly $4.80, which corroborates the card. At 1,000 hours a month that is $4,800 before LLM tokens and before any concurrency limits, which the docs list for TTS and STT but not for Realtime voice agents.
The Ming side has no such number, so any comparison requires an assumption, and the assumption should be visible. inclusionAI's reference 128K serving profile uses tensor parallelism across four 80 GB H800s; long audio inputs can OOM that and force the eight-GPU fallback. At a hypothetical $2 per GPU-hour — a placeholder, not a quote — four cards cost $8 an hour and eight cost $16, both before utilisation, before the idle time that any always-on service carries, and before an engineer to run it. Filled to capacity, the four-GPU configuration is competitive with $4.80 an hour. Filled to ten percent, it is not in the same conversation.
That is the whole cost argument, and it turns on utilisation rather than on either vendor's pricing. A team with steady, high-volume audio traffic and existing GPU capacity can make the arithmetic work. A team testing an idea on Tuesday afternoon cannot, and should not pretend the per-hour figure it computes at 3 a.m. traffic levels means anything.
• Metered — Eleven Realtime v1 Max: $0.08/min, about $4.80/hour, plus LLM usage.
• Self-hosted — Ming-3.0-audio-flash: no rate card; 4 × 80 GB H800 reference, 8-GPU fallback.
• Marginal cost of one more hour — Eleven: $4.80. Ming: effectively zero until the cards saturate, then a step function.
• Cost of the first hour — Eleven: $4.80. Ming: the hardware, the serving stack and the integration.

Languages are where the spec sheets mislead
ElevenLabs says 90+ languages and means it across the TTS and STT stack. Ming-3.0-audio-flash says Chinese and English, and then qualifies it in a way that matters more than the count: "including varied accents and dialects". The vendor's own October table backs that up with dialect error rates that are dramatically better than the omni models it compares against — Hokkien at 12.84 against 28.08 for Qwen3.8-Omni Flash and 81.62 for Gemini 3.8 Flash, Shanghainese at 11.15 against 21.10 and 88.64. If your audio is Mandarin with a regional accent, that is a capability no amount of language-count comparison captures, and Eleven's 90+ list does not tell you how it does on Hokkien.
The reverse is just as true: if your audio is Spanish, Portuguese, Japanese or Arabic, Ming-3.0-audio-flash has no claim at all. This is not a close call on coverage. It is a near-total partition, and it is the single fastest way to decide which column you are in.
Both on one key, for the half that is text
OrcaRouter does not serve Ming-3.0-audio-flash, and it does not serve Eleven Realtime v1 Max either — those are a self-hosted checkpoint and a vendor's own realtime product, and neither is a model we can hand you an endpoint for. What we do serve is the rest of the pipeline either one implies. A voice system built on a self-hosted audio model is mostly text calls: the prompt that tells the audio model what to listen for, the parser that turns its tab-delimited language label into structured output, the summary written from the transcript, the router that sends a hard call to a stronger model. Those run through more than 200 models behind one key, at 0% markup, with automatic failover when a provider degrades — which is what makes it safe to put an unproven audio model behind part of your traffic while the rest keeps working.
Which one to pick
Take Eleven Realtime v1 Max if you want a voice agent this month: 90+ languages, a solved conversational loop, tool calling, a published per-minute price and an OpenAI-compatible session. Accept that its tau-voice task-completion score is 0.578, that its backchannel behaviour is the weakest part of its Full Duplex Bench profile, and that you are renting a stack you cannot inspect.
Take Ming-3.0-audio-flash if your audio is Chinese or English, your hard cases are accents or dialects, your team can run four 80 GB GPUs and a serving stack, and you are willing to build the conversation layer yourself. Accept that the model's own published numbers show it giving up ground on long recordings and speaker diarization, that its time to first audio is twice Eleven's, and that its 99.8 is a leaderboard rank rather than a promise about your workload.

And if the answer is "some of both" — a hosted agent for the languages it covers, a self-hosted model for the accent-heavy traffic — that is not a compromise, it is the shape most production speech systems actually take. The mistake is choosing on the 99s, because at that end of the table the number has stopped telling you anything your own recordings would not tell you faster.
