A generated title card for Realtime 1 vs Realtime Venus contrasting a per-minute hosted price against a per-GPU-hour self-hosted cost, with an Apache 2.0 licence mark, plus an OrcaRouter logo.
Engineering & Research

Realtime 1 vs Realtime Venus: Rent a Voice Agent or Own One for the Cost of a GPU

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Everything about this matchup comes down to a sentence on a licence file. Realtime Venus is Apache 2.0 — a 9B model you download, run on your own A100, and never pay per minute for again. MAI Realtime 1, announced by Microsoft on 17 November 2026, is a hosted API at $0.01 per audio minute with a token-priced reasoning tier behind it and a Foundry preview that bills nothing until public preview. One of these has a price per call. The other has a price per GPU-hour. Comparing them on rate cards is a category error, which is why almost every write-up of hosted-voice-versus-open-weights goes wrong in the same place.

The right question is what each one costs to operate at your volume and what each one can actually do — and there the comparison inverts: the open model's numbers come with a paper describing the harness, while the hosted one's come from a model page with a single accuracy figure on it. That inversion is the story.

What each one actually is

Realtime Venus comes from inclusionAI and is documented in arXiv 2609.13814, credited to Ant Group's Venus Team and Tsinghua University, with the paper dated September 2026. It ships as two checkpoints: Realtime-Venus-Omni, a 9B audio-visual model adapted from MiniCPM-o 4.5 that watches and listens continuously, decides when to speak, and emits text and speech on one timeline; and Realtime-Venus-Audio, a 9B audio-only model on the same streaming backbone. Both use Whisper-Medium as an audio encoder, Qwen3-8B as the language backbone, discrete S3 speech tokens with a streaming flow-matching decoder, a 40,960-token context, BF16 weights, and English and Chinese as their languages.

MAI Realtime 1 is Microsoft's first speech-to-speech model, a hosted two-tier system: 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 route into. No architecture details, no parameter count, no language list. Its one published accuracy figure is 98.8% on Big Bench Audio, on its own model page, vendor-run and unreproduced — which is the mirror image of the situation below, where a smaller lab put its evaluation method in a paper anyone can read.

• Weights — Realtime Venus Apache 2.0, downloadable, two 9B checkpoints; MAI Realtime 1 closed, preview access by registration.

• Languages — Realtime Venus English and Chinese, explicitly; MAI Realtime 1 unlisted.

• Context — Realtime Venus 40,960 tokens; MAI Realtime 1 unpublished.

• Hardware — Realtime Venus needs an A100 for the real-time duplex demo; MAI Realtime 1 needs nothing but a key.

• Modalities — Realtime Venus Omni takes streaming video and audio; MAI Realtime 1 is speech-to-speech.

• Price shape — Realtime Venus per GPU-hour, forever; MAI Realtime 1 per audio minute plus reasoning tokens.

The numbers Realtime Venus published, and what they mean

The paper is unusually specific about full-duplex behaviour, which is the part most voice models leave vague:

• Interruption response — 75% of user interruptions answered, on Full-Duplex-Bench v1.5.

• Continuation after backchannels — 97%.

• Continuation when speech is directed elsewhere — 88%.

• Continuation under background speech — 86%.

• The authors' claim — these "exceed G​emini 3.1 Live and GPT-4o on all three continuation metrics." Note the comparison targets: G​emini 3.1 Live, which G​oogle now marks legacy in favour of 3.8 Live, and GPT-4o. Neither is the current frontier, so the win is real but the bar is a generation old.

On understanding, Realtime-Venus-Audio leads its comparison set on MMAU at 78.0%, MMAU-Pro at 63.2%, Llama Questions at 83.8%, and Speech CMMLU at 67.8%, and ties the top VoiceBench AlpacaEval score at 4.81. Realtime-Venus-Omni is best among evaluated online models on six of eight video benchmarks, including StreamingBench at 70.2%, OVO-Bench at 64.7% and Daily-Omni at 81.3%.

All of those are the authors' own evaluations in their own paper. They are stronger evidence than a marketing page — the harness and comparison set are at least described — and weaker evidence than an independent board. Artificial Analysis does not carry Realtime Venus as of 18 November, so no third party has re-run them.

The one number that is missing from the paper's abstract-level summary is the one that matters for a support agent: a τ-Voice agentic task-completion score. Realtime Venus is built around a different idea — a <delegate> mechanism that pushes asynchronous external tasks to a separate Harness runtime, so the conversation keeps running while work happens in the background. That is architecturally close to what MAI Realtime 1 does internally with its reasoning tier, and to what G​PT Live 1 does by handing off to a backend agent. Three different products, one convergent design: talk fast, delegate the thinking. Nobody has published a head-to-head on the delegation half.

A generated two-column scoreboard titled 'MAI Realtime 1 vs Realtime Venus — the scoreboard'. Left column MAI Realtime 1: licence closed preview only, weights hosted only, price shape $0.01 a minute, languages not published, video no. Right column Realtime Venus: licence Apache 2.0, weights two 9B checkpoints, price shape per GPU hour, languages English and Chinese, video yes, Omni. Footer reads 'Venus figures from its own paper, author evaluations; MAI figures per Microsoft.'

The arithmetic that a per-minute price hides

Here is where the comparison stops being about rates. MAI Realtime 1 at a cent a minute is $0.60 an hour of audio. A single A100 running Realtime Venus costs somewhere in the region of $1.50 to $2.50 an hour on commodity cloud pricing — call it $2 for the sake of argument, and note that this is a market rate, not something either vendor publishes.

• Breakeven — roughly three to four concurrent conversations fill one A100 at those rates. Below that, renting is cheaper. Above it, owning is.

• The catch on the hosted side — MAI Realtime 1's reasoning tier is billed per token on top of the minute rate, so its true cost per hour is above $0.60 and unknown.

• The catch on the self-hosted side — concurrency per A100 for a 9B duplex model is not published either. The demo needs "at least one NVIDIA A100," which tells you it runs and nothing about how many streams it sustains.

So both sides of the breakeven are estimates. What is not an estimate is the shape of the curve: hosted cost scales linearly and forever; self-hosted cost has a fixed floor and then bends. A pilot of a hundred calls a day should rent. A contact centre running two hundred concurrent lines should not, unless the hosted model is materially better — and on this matchup, the open model is the one whose accuracy claims come with a published harness.

There is also an operational asymmetry that never appears on a rate card. Self-hosting a duplex 9B model means owning GPU capacity, cold starts, model loading, and the failure mode where a node goes down and calls drop. Hosted means owning none of that and accepting a vendor's latency, availability and deprecation schedule. Teams consistently underestimate the first and overestimate the second, usually because the first is invisible in a spreadsheet.

Where the hosted model has no open-weight answer

Three things, and they are not small.

Enterprise plumbing: MAI Realtime 1 is coming to Dynamics 365 Contact Center and to Dragon Copilot for nurse-training simulations, with Agoda and Reservations.ai testing. Those are integrations into systems a self-hosted model cannot reach without you building the connectors yourself.

Compatibility: it speaks the O​penAI Realtime API, so an existing voice stack needs a model string rather than a rewrite. Realtime Venus ships as a Python package — AutoModel.from_pretrained, as_duplex(), per-second streaming_prefill() and streaming_generate(), FFmpeg and CUDA on the host. That is a real engineering project.

And transcription breadth: Microsoft's companion announcement, Dragon Copilot Transcribe on MAI Transcribe 2, covers clinical dictation at a scale nothing in the Venus paper addresses — 10 million patient encounters a month, half the multilingual error rate, by Microsoft's own account. If your product is clinical, that ecosystem is the argument, not the per-minute rate.

Where the open weights have no hosted answer

Video. Realtime-Venus-Omni is a streaming audio-visual model — it watches and listens, and the Omni benchmarks are video benchmarks. MAI Realtime 1 is speech. If the interaction involves a camera, a screen share, or a patient gesturing at something, the comparison is not close because the hosted model does not attempt it.

Long-video memory, too: the Venus card claims memory over long video with no extra training, via a use_memory() switch. Nothing in Microsoft's announcement addresses memory across a session.

And the licence. Apache 2.0 means you can fine-tune it, ship it inside a product, run it in a jurisdiction Microsoft has no region for, and keep audio on your own hardware. For a clinical or public-sector deployment where the data cannot leave the building, that is not a preference — it is the requirement, and no amount of per-minute discount substitutes for it.

A screenshot of the Hugging Face model card for inclusionAI/Realtime-Venus captured 18 November 2026, showing the Any-to-Any pipeline tag, Transformers, ONNX and Safetensors badges, English and Chinese language tags, and the multimodal, audio, video, streaming and full-duplex tags.

The hybrid that most teams should actually build

The convergence noted earlier is the practical answer. All three of these products — Realtime Venus, MAI Realtime 1, G​PT Live 1 — separate the conversation from the work: Venus through <delegate> and a Harness runtime, Microsoft through an internal reasoning tier, O​penAI through an explicitly attached backend agent. In every case the hard part is a second model doing the thinking, and in every case that layer is where cost and latency concentrate.

Which means the decision between them is really a decision about the front end, and the backend can be the same either way. OrcaRouter is built for that second layer: 200+ models on one key at provider list price with 0% markup, so whichever front end you pick, the reasoning models behind it are swappable and a vendor price change lands the same day. Automatic failover is what turns a delegated task into something safe to run in production — a background call that fails silently is a promise to a caller that never gets kept, and a routed retry across providers is the difference between a task that completes and one that quietly does not. The routing DSL is how the delegation policy becomes a rule: which intents escalate, to which model, at what budget.

Neither of these two voice models is in our catalogue — Realtime Venus is Apache 2.0 weights you host yourself, and MAI Realtime 1 is MAI Labs and the Foundry limited-access preview. What we route is everything they delegate to.

A screenshot of the OrcaRouter models catalogue captured 18 November 2026, showing the header line '226 models · 17 providers · one API key, one bill' and the model grid with search and sort controls.

Choosing, concretely

Start with Realtime Venus if any of these is true: the data cannot leave your infrastructure, the interaction involves video, you need English and Chinese specifically and nothing else, you have GPU capacity already, or your volume is high enough that a linear per-minute bill is the wrong shape. Accept in exchange that you are running a research-grade stack with a Python API, a paper as your documentation, and no independent benchmark to confirm the authors' numbers.

Start with MAI Realtime 1 if the interaction is speech-only, the volume is modest, the integration is OpenAI-Realtime-shaped already, and a cent a minute is the number that makes the project viable at all. Accept in exchange that you can measure almost none of it: one vendor-reported accuracy figure, a P50 latency with no distribution, no τ-Voice score, no language list, and a two-minute non-commercial demo that was reporting itself unavailable when we checked.

And watch the thing both are quietly betting on. Realtime Venus, MAI Realtime 1 and G​PT Live 1 have all concluded that a voice model should not think hard while someone is waiting. The difference between them is only where the thinking goes. Whoever publishes a task-completion rate for the delegated half first will settle a matchup that the per-minute rates never could.