Realtime-Venus vs Grok Voice Think Fast 2.0: Only One of These Has a Price

Put Realtime-Venus and Grok Voice Think Fast 2.0 next to each other and the first difference is not a benchmark, it is a purchase order. Grok Voice Think Fast 2.0 is a hosted speech-to-speech model that went on sale on 29 July 2026 at $0.08 per audio minute with a published time-to-first-audio of 0.70 seconds and benchmark scores from a third party. Realtime-Venus is a 9B full-duplex system from Ant Group's Venus Team and Tsinghua University that exists as Apache-2.0 weights, a technical report dated 12 September 2026, and nothing you can call. One of these you can wire into a phone line before the end of the week. The other one you can read. That asymmetry turns out to be a much more useful comparison than a scoreboard would be, because the two models made almost the same architectural bet and then diverged completely on what to do with it.

The short answer

Choose Grok Voice Think Fast 2.0 if you need a working voice agent now, in production, with a rate card you can put in a budget and a latency guarantee you can test. Choose Realtime-Venus if you need to own the stack — if your data cannot leave your infrastructure, if you need to fine-tune the front end, or if you are evaluating the architecture rather than shipping a product. There is no third case where they compete, because one has an API and the other does not.

They agree about the hard problem

The interesting thing is how similar the diagnosis is. Both models exist because of the same contradiction: conversation control has to run at sub-second granularity, and reasoning takes seconds. A model that stops to think stops feeling present.

Grok Voice Think Fast 2.0 solves it by reasoning while it speaks. The Think Fast line was built on the idea that the model should not infer first and generate speech second; instead it streams speech and thinks in parallel, so a tool call can fire before the agent has finished its first sentence. xAI reports that version 2.0 uses roughly 0.4 times the reasoning tokens of its predecessor, which is presented as a cost and latency improvement rather than a quality one.

Realtime-Venus solves it by not solving it in the frontend at all. Its dual-loop runtime keeps a one-second interaction loop running — perception, turn control, speech generation — and hands anything expensive to a separate component called Realtime-Venus-Harness through asynchronous delegation markers emitted in the model's own hidden sequence. The foreground conversation never pauses. Background results get reintegrated when they arrive.

Those are different engineering answers to the same question, and they have different failure modes. Reasoning-while-speaking puts the burden on the model to keep both threads coherent. Delegation puts the burden on the runtime to keep the two loops from corrupting each other — which is why the Realtime-Venus paper's causal task capture, an isolated state snapshot per delegated task, is the detail that carries the design.

The one metric they can both be measured on

Full-duplex benchmarks are the only place these two overlap, and they do not overlap cleanly.

• Interruption handling — Realtime-Venus reports responding to 75% of user interruptions on Full-Duplex-Bench v1.5. Grok Voice Think Fast 2.0 scores 95.1% on Full Duplex Bench per Artificial Analysis, up from 77.8% for version 1.0.

• Continuation under overlap — Realtime-Venus reports 97% continuation under backchannels, 88% under other-directed speech, and 86% under background speech, and states it exceeds Gemini 3.1 Live and GPT-4o on all three.

• Different instruments, different authors — the Realtime-Venus figures come from its own technical report with no external reproduction. The Grok figures come from a third party that ran the model. Comparing a self-reported number to an independently measured one is not a comparison, and the gap above is as likely to be methodological as real.

The honest read: on the evidence available, Grok Voice Think Fast 2.0 is the only one of the two whose full-duplex behaviour has been measured by someone with no stake in the result.

Where the independent numbers put Grok Voice Think Fast 2.0

The cleanest current reading comes from Artificial Analysis's Speech Agent Arena, which scores models by blind preference across live voice conversations on tasks like booking a dental appointment and ordering takeout. As of 18 September 2026, Grok Voice Think Fast 2.0 High sits seventh of more than twenty entries at an Elo of 1,011 over 552 samples — behind Google's Gemini 3.1 Flash Live Minimal at 1,096, Gemini 3.8 Live at 1,083 and GPT-Live-1 at 1,053, and well clear of its own predecessor, Grok Voice Think Fast 1.0, at 908.

The column next to the Elo is the more interesting one. On task success rate, Grok Voice Think Fast 2.0 records 94.6%, the highest figure anywhere in the visible top twenty — above Gemini 3.8 Live's 93.2%, GPT-Realtime-2.1 High's 91.5% and GPT-Live-1's 90.9%. Seventh by preference, first by task completion, is an unusual and quite specific profile: listeners do not love the voice, but it finishes the job. If your product does things rather than chats, that is the column that predicts your outcome, and it is the strongest independent evidence either of these two models has.

The figures xAI published at launch are a different instrument and should be read separately: 82.9% on Artificial Analysis's speech-to-speech quality index, first place on the τ-Voice agentic benchmark at 56.5%, 97.2% on Big Bench Audio and 95.1% on Full Duplex Bench. Those were the standings when the model shipped in late July 2026, and boards in this category move monthly — which is exactly why the arena table above, read today, is worth more than the launch numbers.

The bill, and the parts of it that are not on the bill

Grok Voice Think Fast 2.0 costs $0.08 per minute of audio, roughly $4.80 an hour, plus $0.004 for each text input message. That is a 60% increase over the $0.05 per minute the 1.0 version charged at launch, and xAI moved the rolling grok-voice-latest alias onto 2.0 automatically on 5 August 2026, so teams that wanted to stay on the old price had to pin an explicit version before that date. The per-minute model also means efficiency improvements accrue to the vendor: if the model gets better at finishing calls faster, your bill per call falls, but your cost per minute does not.

Realtime-Venus has no bill, because it has no service. What it has instead is a hardware floor. Two 9B checkpoints in BF16 are roughly eighteen gigabytes of weights each before activation memory, and the card documents a per-second streaming loop that has to run continuously. A GPU node capable of that has a monthly cost, and it is a fixed one — you pay it whether or not anyone calls. For a product with steady traffic, that is cheaper than $4.80 an hour. For a product with intermittent traffic, it is dramatically more expensive, and the crossover point is the only financial question that matters here.

Weights, control, and what each one lets you change

This is where the two models stop being comparable at all.

• Fine-tuning — Realtime-Venus ships weights. You can fine-tune them, quantise them, run them air-gapped, and inspect every layer. Grok Voice Think Fast 2.0 is a closed hosted model; what you can change is the prompt, the voice selection, and the tools you register.

• Voice — Grok Voice Think Fast 2.0 ships preset voices and supports custom voice cloning from roughly one minute of speech. Realtime-Venus ships a reference voice and a bundled token-to-waveform decoder, and anything beyond that is your problem.

• Reach — Grok Voice Think Fast 2.0 supports 25-plus languages and speaks PCM16 at 24 kHz by default with μ-law for telephony, over a WebSocket session that is OpenAI Realtime-compatible. That compatibility is a real migration asset: most existing realtime voice code needs a base URL and a key change. Realtime-Venus documents English and Chinese.

• Video — Realtime-Venus-Omni takes streaming video and images through a SigLIP2 encoder and does continuous visual perception. Grok Voice Think Fast 2.0 is audio-and-text; there is no camera path.

• Long context — Realtime-Venus carries a 40,960-token context and a training-free long-video memory that can be enabled on a forty-minute window. Grok Voice Think Fast 2.0 is a session model with no comparable published memory feature.

Where each one breaks

The failures worth planning for are opposite. Grok Voice Think Fast 2.0's exposure is vendor concentration and unverifiable marketing: xAI's Starlink A/B results, its transcription-accuracy multipliers, and its speed framing are all company-reported with no published underlying data, and a per-minute model that changes price by 60% between versions has demonstrated it will change price. Pin your version and run your own evals on your own audio.

Realtime-Venus's exposure is that nobody has run it. Its benchmarks are self-reported, its harness is undocumented relative to its importance, and its latency in a real deployment is unknown — the report describes a one-second interaction cadence, which is a design parameter, not a measured time-to-first-audio. A model with no API and no reproduction has no track record, only a specification.

There is a middle path that is worth stating plainly, because it is what most teams will actually do. If you are building a delegation-based system — pick your own front end, hand the expensive work to a text model — the front end and the reasoning leg do not have to come from the same place. OrcaRouter carries neither Realtime-Venus nor Grok Voice Think Fast 2.0; both voice layers sit outside what we serve, and we say so rather than implying otherwise. The delegated leg is a different matter. Nearly 200 text models sit behind one key at provider list price with no markup, with automatic failover so an upstream outage does not drop a live call, a routing DSL for composing several models into one call, and model fusion for decisions that deserve more than one opinion. That is the half of a voice agent that benefits from being swappable, and it is the half you can change without touching the model that is holding the conversation.

Which one to pick

Pick Grok Voice Think Fast 2.0 this quarter. It is available, it is independently benchmarked, it is first on the agentic evaluation that predicts whether your agent can do anything useful, and 0.70 seconds to first audio is a number you can build a user experience around. Budget for the 60% price rise over 1.0 and pin your model version.

Pick Realtime-Venus only if the constraint is ownership. If your deployment cannot call an external API, if you need to fine-tune the conversational front end, or if you need the camera — those three cases are the entire case, and they are strong ones when they apply.

What should not decide it is the benchmark table, because the two sides of it were produced by different people under different conditions. Grok Voice Think Fast 2.0's numbers were measured by someone else. Realtime-Venus's were not. Until that changes, the comparison that is actually available is between a product and a paper.