Higgs Realtime Duplex Explained: Boson AI’s Full-Duplex Model That Talks While It Works

Higgs Realtime Duplex Explained: Boson AI’s Full-Duplex Model That Talks While It Works

Higgs Realtime Duplex is Boson AI’s first full-duplex interactive omni model, introduced on November 5, 2026 as a research preview. Unlike the voice assistants most of us know — which wait for you to finish, think, and then reply — Higgs Realtime Duplex listens, watches, and talks at the same time. It even renders a face of its own in real time, in sync with its voice, while it keeps thinking, calling tools, and working in the background without ever pausing the conversation.

If you have ever talked to a voice assistant and felt the awkward stop-start rhythm — you speak, it waits, it thinks, it answers — you already understand the problem Boson AI is trying to solve. Human conversation does not work in turns. People interrupt each other, say “mm-hm” while the other person is still talking, answer before a question is finished, and stay quiet when a remark was not meant for them. Higgs Realtime Duplex is built to do all of that, and to keep working on your task while the conversation flows. In this guide, we explain what Higgs Realtime Duplex is, how its 160-millisecond chunking works, what its benchmarks actually say, how it compares with OpenAI’s realtime models and Google’s Gemini Live, what it costs to run, and how you can join the API waitlist today.

Quick note before we start: the Duplex API is not open yet. It is a research preview with a waitlist, and Boson AI has published no price for it. The earlier turn-based Higgs Realtime model is available today. We flag every claim that comes from Boson AI’s own testing, so you can tell the difference between a verified fact and a company-reported score.

Table of Contents

What Is Higgs Realtime Duplex?

Higgs Realtime Duplex is a 45-billion-parameter mixture-of-experts (MoE) model from Boson AI that handles live conversation over voice and video in a single model. It watches your camera or shared screen, speaks with a generated voice, and shows a generated face that moves in sync with what it says — all at the same time, with no turn-taking. Boson AI describes it as the next generation of its earlier Higgs Realtime speech-to-speech model, and as its first full-duplex interactive omni model. The word “omni” matters here: the same model takes in audio, video, and text, and produces voice and a face, instead of bolting a speech module onto a text model.

The headline idea is simple to state and hard to build: the model talks while it works. In Boson’s own demos, Higgs Realtime Duplex presents an architecture diagram in a team meeting, takes questions mid-sentence, reads the whiteboard, and updates the diagram as the discussion goes on. In another demo, an engineer reads a book the whole time while the model diffs two engine revisions by voice alone, asks Claude Code to build a missing mass-calculation tool in the background, runs a compliance check, applies the fixes, emails the designer, and reports the new mass once the tool lands. The conversation never pauses for the work, and the work never waits for the conversation.

Three design goals drive the whole project, according to Boson AI. First, it has to feel human: in real conversation, a response typically lands within 200 milliseconds of the other speaker finishing, and people sometimes answer before the other person finishes. Second, it has to be smart: most conversations with an agent are about getting something done, which takes emotional intelligence and real reasoning. Third, it has to cost little: Boson believes talking will become the main way people work with agents for hours every day, so its design target is that an hour with the model costs at most one-tenth as much as an hour of a person’s time. Keep those three goals in mind, because they explain nearly every engineering choice that follows — from the chunk size to the token budget to the size of the model itself.

Higgs Realtime Duplex voice assistant - woman talking to an AI voice assistant at home

Boson AI is a smaller lab than OpenAI or Google, and that context is part of the story. The company trains on its own NVIDIA H100 cluster in a data center it built itself, and its co-founder is Alex Smola, a well-known machine learning researcher. Higgs Realtime Duplex is the company’s bid to compete on interaction quality rather than raw scale: a 45B front model that trades blows with models six times its size by being designed for conversation from the ground up, rather than adapted to it after the fact.

How Full-Duplex Actually Works: 160 ms Chunks

To understand Higgs Realtime Duplex, you first need to understand what “full duplex” means — and why almost every voice agent you have used is only half duplex. In a half-duplex system, the conversation moves in turns: you talk, the system listens; the system talks, you listen. Your phone line, by contrast, has always been full duplex: both people can speak at the same time, interrupt each other, and talk over each other. Higgs Realtime Duplex brings that phone-line behaviour to AI conversation.

The core design idea is to slice time, not turns. Boson cuts every stream — your voice and camera feed coming in, tool results and backend replies coming in, the model’s voice, face, and background work going out — into 160-millisecond chunks. At each step, the model reads the chunk that just arrived and writes the chunk that plays next. Silent intervals are represented as chunks too, and with video enabled, frames keep streaming in even during silence, so the model never stops listening and watching.

That 160 ms number is not arbitrary. It is short enough that the model’s behaviour feels continuous — it can say “mm-hm” while you talk, stop the moment you cut in, and answer before you finish when the timing is right — and long enough that the model can do meaningful work per step. In Boson’s sample conversation, each cell of a grid represents one 160 ms chunk: the model says “mm-hm,” thinks, and calls a tool while you talk; a tool result arrives as you finish; the model replies, then pauses its thinking; when you cut in, it stops, re-plans, and delegates, then resumes talking before the backend returns. Its face renders on every step.

Thinking and tool-call generation can span multiple chunks, and concatenating those chunks recovers the full output stream. Each arriving tool result enters the model’s input chunk as a payload, without a fixed per-chunk text-token quota. Some tools return intermediate results in stages as they work, and the model does not wait for a full result before it starts to speak. That is the detail that makes the conversation feel alive: the model starts answering from partial results, the way a person starts answering a question before they have worked out every detail.

Higgs Realtime Duplex and the 200-Millisecond Rule

Why does any of this matter to a user? Because human conversation has a strict timing budget. Research on conversation shows that a response typically lands within about 200 milliseconds of the other speaker finishing — and Boson cites exactly that figure. A system that waits a full second or two before answering feels broken, even if the answer itself is perfect. Worse, a system that cannot be interrupted forces you to listen to a long answer you no longer need, or to repeat yourself after it finishes. Higgs Realtime Duplex is engineered around that 200 ms budget: the 160 ms chunk is the unit that lets it react inside a single conversational beat.

The model also has to decide when to speak, which is a harder problem than deciding what to say. It says “mm-hm” while you talk, cuts in to help when you get stuck, and stays quiet when a remark was not addressed to it. The drive-thru demo in Boson’s announcement shows why that matters: the model is asked to check the markets and a training run while the driver orders with the cashier. It stays quiet during the order, then reports back once the order is done — and at the pickup window it confirms the driver never ordered animal-style fries, because it heard every word and knew which ones were for it. Ignoring remarks that are not addressed to you sounds trivial until you try to build it; it is one of the hardest behaviours in live voice AI, and it is exactly the kind of thing that separates a demo from a product people can actually use.

One Model, Many Threads: Thinking While Talking

The second big idea in Higgs Realtime Duplex is that the front model — the part that talks with you directly — runs many threads at once. Your voice and video stream in while the model’s voice and face stream out. At the same time, it can think, call a tool, or hand a question to a bigger model, and none of that stops it from listening and watching. The front model thinks on its own and makes async calls straight to the backend model and to tools; their results come back while it keeps talking and listening.

Boson gives the front model three ways to raise the quality of an answer, and each one is designed to run without pausing the conversation:

  • Think. Thinking usually takes seconds, and nobody wants to wait seconds for every answer. As each meaningful phrase or idea arrives, Higgs Realtime Duplex can build on an ongoing reasoning thread without waiting for the full question. That reasoning continues across chunks, alongside both listening and speaking, so it can keep thinking even as it talks.
  • Call tools. Search, knowledge bases, booking, CRM lookups, or any function a deployment defines. Calls go out as early as possible, asynchronously, and the model starts answering from partial results. The front model keeps listening, thinking, and speaking while tools work.
  • Delegate. When a task is beyond the front model, it hands it to a much bigger backend model, along with the full shared context. While the backend works, the front model keeps the conversation going, including on other topics, and brings the result back when it is ready. Boson says this works best when the backend is a frontier model.

This is also where Boson draws its sharpest contrast with Thinking Machines Lab. In Thinking Machines’ interaction models, the front model leaves thinking and tool calls to a background model. Boson’s front model thinks and calls tools itself. Boson’s argument is practical: a system that delegates all reasoning and tool use to a larger model is simpler to train, but a large reasoning model may take ten seconds or more to respond, and calling it every turn makes cost hard to control. So the front model handles most requests on its own and saves the backend for the turns that genuinely need it. Whether that trade-off is right will only become clear once developers can test both systems side by side — but it is a deliberate, clearly stated design choice, not an accident.

How Higgs Realtime Duplex Was Trained

Under the hood, Higgs Realtime Duplex is a 45B mixture-of-experts (MoE) model. Boson says it tried larger models — they do score higher — but they cost more to run, and since the hardest questions can go to a bigger backend model anyway, a mid-sized front model with solid intelligence gives the best balance. Training started from a text-pretrained model and continued on about 2 trillion tokens of text, audio, and visual data. The text training preserves the general intelligence the model learned from language; the audio and visual training teaches it to process both modalities directly and connect what it hears, sees, and says. Training ran on Boson’s own NVIDIA H100 cluster, using a training stack based in part on Megatron.

For post-training, Boson uses multi-teacher on-policy distillation (MOPD). Starting from the same base model, the team trains specialist teachers with reinforcement learning — including GRPO — on groups of related tasks: natural conversation, interactive behaviour, dynamic delegation, and task completion. Through interaction, the teachers learn what to say and when to act: when to speak or yield, when to begin or extend a reasoning thread, when to call a tool, or when to delegate. Those teachers are then distilled into a single student model that combines all of those capabilities.

There is also a recursive self-improvement (RSI) loop: Boson’s agent system interacts with simulated environments, reviews its failures, synthesises targeted data, and trains the model on it before repeating the cycle. The simulation has to reproduce the timing and variety of real interaction — interruptions, overlapping speech, side conversations, background noise — realistic enough to teach natural behaviour and hard enough to expose weaknesses. For a small team, Boson says, the loop turns edge cases into training data and lets the team iterate with far less manual data work. That last point is worth noting: the hardest part of building a full-duplex model is not the architecture, it is the training data for all the messy moments in real conversation, and simulation is how a small lab generates that data at scale.

The Face: Zero-Shot Voice and Face Cloning

One of the most striking parts of Higgs Realtime Duplex is that it has a face — and that the face is not baked into the model. Boson provides default voices and faces, but they are loaded zero-shot at deployment time: a short reference audio clip sets the voice, and a single reference image sets the face, with no fine-tuning. The references go into the two decoders that render the model’s output, not into the model itself, so swapping them changes the voice and the face instantly.

Each output chunk carries two audio frames and two image latents. A causal flow-matching decoder turns the audio frames into speech, conditioned on the reference clip, so the speech matches that voice. A causal DiT — a diffusion decoder — turns the latents into video, conditioned on the reference image and the previous two latents, so the face stays visually consistent and continuous for as long as the conversation runs. The two renderers work from one stream: the model decides what to say and how its face should move, and the decoders turn those decisions into audible speech and visible expression.

Why does a face matter at all? Because conversation is not just words. When you talk to a person, you read their face for timing cues: are they about to speak, are they listening, are they confused, are they done? A voice-only agent gives you none of that, which is one reason voice calls with AI can feel oddly tiring — you are doing all the timing work yourself. A face that moves in sync with the voice restores some of those cues. It also changes what the model can be used for: a tutor that watches your mouth as you practise pronunciation, a meeting participant that presents slides, a character in a game that reacts to what you say. Whether users actually want their AI assistant to have a face is an open question — some will find it helpful, others uncanny — but Boson is betting that the face is part of what makes a live agent feel present rather than remote.

There is an obvious safety question here too. Zero-shot voice and face cloning — a short clip and a single image are all it takes — is exactly the technology behind convincing voice and video impersonation. Boson does not report voice-cloning or avatar-cloning benchmark results, noting that the other models in its comparison do not support those features. The company has not published details of consent checks or watermarking for cloned voices and faces in the Duplex announcement. If you plan to deploy this technology, that is a question to ask before you build: who consented to this voice and this face, and how is that consent checked? We will come back to this in the limitations section.

What Higgs Realtime Duplex Can Do: Real Scenarios

Boson AI’s announcement does not just describe Higgs Realtime Duplex in the abstract — it shows the model working through a set of concrete scenarios, each one chosen to demonstrate a different capability. Taken together, they are the best evidence of what the model is actually for, and they are worth walking through one by one, because each one maps to a real product category.

Higgs Realtime Duplex as a Meeting Assistant

In the headline demo, Higgs Realtime Duplex joins a team meeting, presents an architecture, takes questions mid-sentence, reads the whiteboard, and updates a diagram as the talk goes on. This is the hardest scenario of the set, because it combines everything at once: the model has to watch a shared screen, listen to several people, decide when to speak, answer questions while still presenting, and do useful work — updating a diagram — in the background. If it works as shown, it is a genuine step beyond today’s meeting assistants, which mostly transcribe and summarise after the fact. A live participant that can present and take questions is a different product from a note-taker. Our take: this is the demo to watch most closely, because it is also the one most likely to be flattered by a staged recording. The question is how it handles a real meeting — overlapping speakers, bad audio, a whiteboard nobody can read.

Higgs Realtime Duplex video call - woman on a video call with an AI assistant on a laptop

Higgs Realtime Duplex as a Spoken-English Tutor

The tutor demo is smaller and, in some ways, more convincing. Higgs Realtime Duplex times a picture-description exercise, catches the student saying “west” when they meant “vest,” watches the student’s mouth as they practise, and paces the drill like a real coach. This works because the model’s two superpowers — watching and listening at the same time — are exactly what a pronunciation coach needs. A text chatbot cannot see your mouth. A turn-based voice tutor cannot correct you mid-word without breaking the flow. A full-duplex tutor with a camera can do both, and it can do it at the pace of a real lesson. Language learning is one of the clearest near-term markets for this kind of model, and it is no accident that Boson chose it as a demo: it is a use case where the face, the camera, and the interruption handling all matter at once.

Higgs Realtime Duplex at the Drive-Thru

The drive-thru demo is the most entertaining — and the most technically revealing. The model is asked to check the markets and a training run while the driver orders with the cashier. It stays quiet during the order, then reports back once the order is done. At the pickup window it confirms the driver never ordered animal-style fries: it heard every word, and knew which ones were for it. This is the “ignore remarks not addressed to it” capability in action, in a noisy environment with two conversations happening at once. It is also a genuine commercial use case: drive-thru voice ordering is already a real market, and a model that can tell the customer’s words apart from the cashier’s — and keep working on a background task while it waits — is exactly what a drive-thru operator needs.

Higgs Realtime Duplex drive-thru - car at a fast food drive-thru ordering window

Higgs Realtime Duplex as a Hands-Free Engineer

In the engineering demo, an engineer reads a book the whole time while the model works by voice alone: it diffs two engine revisions, asks Claude Code to build a missing mass-calculation tool in the background, runs a compliance check, applies the fixes, emails the designer, and reports the new mass once the tool lands. This is the “works while you talk” idea at full stretch — not just answering questions, but orchestrating other tools and agents (in this case, Claude Code) while the user does something else entirely. It is also the demo that most clearly shows the delegation architecture: the 45B front model keeps the conversation going while the heavy work happens elsewhere. For developers and engineers, this is the scenario that matters most, because it points at a future where the voice interface is not a toy but the control surface for real multi-agent work.

Higgs Realtime Duplex in Games and Teaching

Two more demos round out the set. In one, a Minecraft villager named Anya haggles with the player over emeralds and bread — a character that listens, watches, and negotiates in real time, with a face of its own. In another, a “Talk to Character” feature answers lecture questions about attention scores and softmax — a tutor that knows the material and can discuss it live. Games and education are natural homes for a model with a face: a character that can see you, hear you, and react is a different experience from a scripted non-player character, and a tutor that can watch you work through a problem is different from a video lecture. These demos are early, but they hint at where the consumer market for Higgs Realtime Duplex may actually start — not in enterprise contact centres, but in products where a live, visible character is the point.

Higgs Realtime Duplex voice conversation - woman speaking to a smart speaker voice assistant

Higgs Realtime Duplex Benchmarks Explained

Boson AI published a benchmark comparison of Higgs Realtime Duplex against the Thinking Machines Lab interaction model and the realtime and live models from OpenAI and Google. The suite covers spoken question-answering, full-duplex interaction, joint video and audio understanding, proactive video (the model watches a clip and chooses when to speak), agentic task completion, and text instruction following. All of these figures are company-reported — Boson says it reproduced more than half of the competitor scores to check its evaluation pipeline, but no independent lab has verified the results, so treat them as Boson’s own claims, not settled facts.

Benchmark Higgs Realtime Duplex (45B MoE) TML Interaction Small (276B MoE) GPT Realtime & Live series Gemini Live series
Big Bench Audio accuracy 98.9 (highest) 96.5 RT-2 Minimal 72; RT-2 High 97; GPT-Live-1 / Astra medium 90 3.1 Flash Live Minimal 71; 3.1 Flash Live High 97; 3.8 Live Ext. Thinking High 98
Audio MultiChallenge APR 44.9 (2nd highest) 43.4 RT-2 Minimal 37.6; RT-2 xHigh 48.5 3.1 Flash Live Minimal 26.8; 3.1 Flash Live High 36.1
Full-Duplex-Bench v1.5 average 75.5 77.8 RT-2 Minimal 46.8; RT-2 xHigh 47.8; GPT-Live-1 80.1 3.1 Flash Live Minimal 54.3; 3.1 Flash Live High 45.5
QIVD streaming accuracy 56.0 54.0 RT-2 Minimal 57.5; RT-2 xHigh 58.2 3.1 Flash Live Minimal 54.7; 3.1 Flash Live High 56.1
ProactiveVideoQA PAUC@ω=0.5 31.9 (2nd highest) 33.5 RT-2 Minimal 25.0 —
τ-Voice task completion (pass@1, %) 69.8 (2nd highest) — RT-2 Minimal 30.8; RT-2 High 39.8; GPT-Live-1 / Astra medium 74.5 3.1 Flash Live Minimal 26.2; 3.1 Flash Live High 37.7; 3.8 Live Ext. Thinking High 68.6
IFEval accuracy 94.1 (2nd highest) 89.7 RT-2 Minimal 89.6; RT-2 xHigh 95.2 3.1 Flash Live Minimal 85.8; 3.1 Flash Live High 90.0

Two things stand out in that table, and Boson itself points to both. First, GPT and Gemini models usually reach their best scores with thinking set high — which costs latency and money in a live conversation, so a headline benchmark number and a usable live number are not the same thing. Second, model size is a real distinction: the Higgs Realtime Duplex front model has 45B parameters, compared with 276B for TML Interaction Small — about six times as many. Even so, the results trade blows across the suite. Boson reads this as a sign that the field is still early, and that data, architecture, and training methods matter more than raw scale. That is a fair reading of its own table, but it is also the reading a smaller lab would give: the honest test is whether the gap holds up when independent evaluators run the same benchmarks, and when the models are tested in live deployments rather than on a benchmark harness.

It is also worth noting what the table does not show. Higgs Realtime Duplex does not lead on Full-Duplex-Bench v1.5 — the benchmark most directly about the thing it claims to do best — where GPT-Live-1 scores 80.1 against its 75.5. On τ-Voice task completion, GPT-Live-1 with Astra medium leads at 74.5% against 69.8%. On IFEval, GPT’s RT-2 xHigh leads at 95.2 against 94.1. In other words, Higgs Realtime Duplex is competitive across the board and best-in-class on one benchmark, but it is not dominant. That is a perfectly respectable result for a 45B model from a small lab — and it is a much more useful picture than the headline “98.9, highest score” alone.

Higgs Realtime Duplex vs GPT Realtime, Gemini Live and Thinking Machines

The full-duplex voice market moved fast in the first two weeks of November 2026. GPT-Live-1 arrived on Retell on November 2, Soniox shipped STT v6 on November 3, and Higgs Realtime Duplex followed on November 5. A recent analysis of the fortnight puts all four launches in one frame: every one of them bets on listening while talking, and on measuring barge-in — how cleanly a model handles being interrupted — on real recordings. That is the right frame, because the differences between these models are not about who can answer a quiz question. They are about who behaves best when a real person interrupts, talks over them, or changes their mind mid-sentence.

Here is how Higgs Realtime Duplex stacks up against the main alternatives on the dimensions that matter for a live deployment:

  • OpenAI GPT Realtime and GPT-Live-1. OpenAI’s realtime models are the incumbents, with the widest deployment and the strongest developer ecosystem. On Boson’s own table, GPT-Live-1 leads on Full-Duplex-Bench (80.1 vs 75.5) and on τ-Voice task completion (74.5% vs 69.8%), while Higgs Realtime Duplex leads on Big Bench Audio (98.9 vs 90). GPT’s models reach their best scores with thinking set high, which costs latency; Higgs Realtime Duplex is designed to think continuously in small slices instead. OpenAI also has one advantage Boson cannot match yet: availability. GPT Realtime is a shipping product with published pricing; Higgs Realtime Duplex is a research preview on a waitlist.
  • Google Gemini Live. Gemini Live is the most widely distributed live model in the world, because it ships on Android phones and inside Google’s consumer products. On Boson’s table, Gemini Live’s best configurations are close to Higgs Realtime Duplex on Big Bench Audio (98 on 3.8 Live Ext. Thinking High) but far behind on Audio MultiChallenge and τ-Voice task completion. Gemini Live’s strength is distribution and multimodal grounding inside Google’s ecosystem; its weakness, for developers, is that it is primarily a consumer product, not a platform you build on. If you are building a product, Higgs Realtime Duplex and GPT Realtime are the comparison that matters. If you are a user, Gemini Live is the one already in your pocket.
  • Thinking Machines Lab interaction models. TML’s interaction model is the closest architectural cousin: a large front model (276B MoE) built for live interaction. The key difference is philosophical. TML’s front model leaves thinking and tool calls to a background model; Boson’s front model thinks and calls tools itself. On the benchmarks, the two trade blows — TML leads on Full-Duplex-Bench (77.8 vs 75.5) and ProactiveVideoQA (33.5 vs 31.9), while Higgs Realtime Duplex leads on Big Bench Audio (98.9 vs 96.5) and IFEval (94.1 vs 89.7). The real difference is cost: a 45B front model is far cheaper to run than a 276B one, and Boson is betting that the smaller model plus a backend for hard questions beats one giant model for every turn.
  • Phonic-1 and other voice-agent models. The same fortnight also brought Phonic-1, which we covered on this site on November 1, and GPT-Live-1 on Retell. Phonic-1 is a voice-agent model built for phone calls, where the hard problems are latency, interruption handling, and background noise — the same problems Higgs Realtime Duplex is built for. The difference is scope: Phonic-1 is voice-only, while Higgs Realtime Duplex adds video in and a face out. If your product is a phone agent, voice-only may be all you need. If your product is a tutor, a meeting participant, or a game character, the face and the camera change what you can build. For the broader voice-agent landscape, our Phonic-1 explainer is a useful companion read.

The honest summary is that there is no clear winner yet, and that is exactly what makes this a moment worth watching. Four credible models arrived within two weeks, each with a different bet: OpenAI on ecosystem, Google on distribution, Thinking Machines on scale, Boson on interaction design and cost. The benchmarks are close enough that the deciding factor will be live behaviour — latency, interruption handling, and cost per hour — not a single leaderboard number.

What Does Higgs Realtime Duplex Cost to Run?

Cost is one of the three design goals Boson states for Higgs Realtime Duplex, and it is the one with the most concrete numbers attached. In full duplex, the model processes every chunk even when nobody is talking, so context fills fast. Boson squeezes every stream to keep the token budget tight: one audio token per chunk in, two out, a camera frame only every six chunks, and two latents for the face. Audio in runs at 6.25 tokens per second and audio out at 12.5; a camera frame, about a dozen tokens, comes once every six chunks.

The result is that the audio and visual streams use about 2,600 tokens per minute. With optional thinking, tool calls, and their results, the total is roughly 2,600 to 5,000 tokens per minute, depending on the workload. That estimate already assumes progressive compaction of intermediate reasoning and tool traces, with the information needed for ongoing work carried forward. In production, the model usually retains the last five minutes of audio and visual streams alongside this compacted text — roughly 13,000 to 25,000 tokens for the recent window, depending on activity and compaction. Older history is retained as text on top of that budget.

What does that mean in dollars? Boson has not published a price for the Duplex API, so there is no per-minute dollar figure yet — and any article that gives you one is guessing. What we can say is what Boson’s design target implies: an hour with the model should cost at most one-tenth as much as an hour of a person’s time. If a person’s hour costs, say, $30, the target is $3 per hour or less; if it costs $100, the target is $10 or less. At 2,600 to 5,000 tokens per minute, an hour is 156,000 to 300,000 tokens — so the implied price band depends entirely on the per-token rate Boson eventually sets. The token budget is genuinely tight for a multimodal model, and it is the strongest evidence that Boson is serious about the cost goal. But until the pricing page for Duplex goes live, the honest answer to “what does it cost?” is: we know the token budget, we know the target, and we do not know the price.

For comparison, the existing Higgs Realtime model — the turn-based version available today — does have published pricing on Boson’s pricing page. If you need to budget a voice-agent project right now, that is the number to look at, and it is the practical reason to start with Higgs Realtime today and move to Duplex when the API opens. We link to Boson’s official pricing page in the sources below rather than quoting a figure that may change.

Higgs Realtime vs Higgs Realtime Duplex: What Is Available Today?

The naming is confusing, so let us be precise. There are two products with similar names, and only one of them is new:

  • Higgs Realtime is Boson’s existing speech-to-speech model, available today. Its documentation describes a full-duplex voice API over WebSocket with turn detection, interruptions, tool calling, and an OpenAI-compatible event protocol. It has published pricing. It is the model you can actually build with right now.
  • Higgs Realtime Duplex is the new model announced on November 5, 2026. It adds continuous video in, a face out, and reasoning that runs alongside the conversation — the front model thinking and calling tools itself instead of handing that work off. Its API is “coming soon” via a waitlist. There is no public price and no open weights announced.

In short: Higgs Realtime is the product, Higgs Realtime Duplex is the research preview of the next generation. That distinction matters for two reasons. First, if you read coverage that says “Higgs Realtime Duplex is available now,” it is wrong — check whether the writer means the older model. Second, if you want to try Boson’s technology today, you do not need to wait: the Higgs Realtime API is live, and building with it now is the fastest way to be ready when Duplex opens. Boson’s own announcement makes exactly this point, directing readers to the existing Higgs Realtime documentation while the Duplex API is prepared.

It is also worth noting that Boson is not the only lab shipping in this space right now. Google launched Gemini 4 Hydrogen on November 12, 2026 — a frontier model whose live and realtime capabilities are part of the same competitive race — and the agent-infrastructure cluster around it, from Perplexity Teammates to Salesforce’s Agentforce coworker in Microsoft Teams, shows how fast the whole category is moving. Higgs Realtime Duplex is one launch in a very crowded fortnight, and its fate will depend as much on how fast Boson ships the API as on how good the benchmarks look.

How to Join the Higgs Realtime Duplex Waitlist

As of November 13, 2026, there is no public API, no playground, and no download for Higgs Realtime Duplex. The only official route is the waitlist, and here is how it works:

  1. Go to Boson AI’s official announcement at boson.ai/blog/higgs-realtime-duplex and read the full post, including the benchmark table and the demo descriptions.
  2. At the end of the post, click the join the API waitlist link. That is the official waitlist — do not trust third-party pages claiming to offer early access, downloads, or API keys.
  3. While you wait, try the existing Higgs Realtime model, which is available today with documentation at docs.boson.ai/models/higgs-realtime and published pricing on Boson’s pricing page. Building a small voice agent on Higgs Realtime now teaches you the event protocol, the tool-calling pattern, and the interruption behaviour you will need when Duplex opens.
  4. Watch Boson AI’s official channels — its blog, its LinkedIn page, and its docs — for the API launch announcement. Research previews from small labs can stay in waitlist for weeks or months, so do not plan a product launch around a date Boson has not given.

A word of caution: whenever a new AI model launches, fake download pages, fake API keys, and fake “early access” offers appear within days. Boson AI has announced no price, no open weights, and no public download for Higgs Realtime Duplex. If a site offers you a download or an API key, it is not from Boson. The waitlist link in the official announcement is the only legitimate route.

Limitations and Honest Caveats

A good explainer says what a product cannot do as clearly as what it can. Here is the honest list for Higgs Realtime Duplex:

  • It is a research preview, not a product. The Duplex API is on a waitlist with no launch date, no price, and no service guarantees. Everything in this article about using it is forward-looking. If you need a full-duplex voice model today, use Higgs Realtime, GPT Realtime, or Gemini Live — all shipping products.
  • All benchmarks are company-reported. Boson says it reproduced more than half of the competitor scores, but no independent evaluator has confirmed the table. The Big Bench Audio lead of 98.9 is the headline; the Full-Duplex-Bench result of 75.5, behind GPT-Live-1’s 80.1, is the number that matters most for the model’s core claim — and Boson does not lead there.
  • It does not lead on every benchmark. As the table above shows, Higgs Realtime Duplex is second or third on several measures, including task completion and instruction following. It is competitive, not dominant — a strong result for a 45B model, but not a reason to switch from a shipping product on benchmarks alone.
  • The demos are staged. Every demo in the announcement is a recording produced by Boson. The meeting demo, the tutor demo, and the drive-thru demo all show the model working well. None of them shows it failing, misunderstanding, or handling a bad connection. Real deployments will be messier, and the model’s behaviour in that mess is untested in public.
  • Zero-shot voice and face cloning raises safety questions. A short audio clip and a single image are all it takes to set the voice and face. That is powerful for legitimate products — a tutor with a consistent character, a brand voice — and dangerous for impersonation. Boson has not published consent or watermarking details for Duplex, and it does not report voice-cloning or avatar-cloning benchmark results. Ask about safeguards before you deploy.
  • It is a small lab competing with giants. Boson AI trains on its own H100 cluster and publishes impressive results, but OpenAI and Google have vastly larger compute, distribution, and developer ecosystems. The history of AI is full of technically excellent models from small labs that lost because the bigger lab shipped faster and integrated better. Boson’s bet is that interaction quality wins; the market will decide.
  • The cost target is a target, not a price. “An hour costs at most one-tenth of a person’s hour” is a design goal. Until Boson publishes per-token pricing for the Duplex API, you cannot budget a deployment on it.

Frequently Asked Questions

What is Higgs Realtime Duplex?

Higgs Realtime Duplex is Boson AI’s first full-duplex interactive omni model, announced on November 5, 2026 as a research preview. It listens, watches, and talks at the same time, renders a face of its own in real time in sync with its voice, and keeps thinking, calling tools, and delegating to a larger backend model in the background — all without pausing the conversation. The front model is a 45B mixture-of-experts model trained on about 2 trillion tokens of text, audio, and video.

Is Higgs Realtime Duplex available now?

No. As of November 13, 2026, the Duplex API is “coming soon” via a waitlist. The earlier Higgs Realtime model is available today, with documentation and published pricing on Boson’s site. If you want to build with Boson’s technology now, start with Higgs Realtime.

How much does Higgs Realtime Duplex cost?

Boson has not published a price for the Duplex API. What it has published is a token budget — about 2,600 tokens per minute for audio and video, or roughly 2,600 to 5,000 per minute with thinking and tool calls — and a design target: an hour with the model should cost at most one-tenth as much as an hour of a person’s time. The dollar price will depend on the per-token rate Boson sets at launch.

What does “full duplex” mean in Higgs Realtime Duplex?

Full duplex means input and output flow at the same time, the way they do in a phone call. Higgs Realtime Duplex cuts every stream into 160-millisecond chunks: your voice and video come in while its voice and face go out, and it can say “mm-hm,” answer before you finish, stop when you cut in, and ignore remarks that are not addressed to it. Most voice assistants are half duplex — they take turns. Full duplex is what makes a conversation feel human.

How does Higgs Realtime Duplex compare with GPT Realtime and Gemini Live?

On Boson’s own company-reported benchmarks, Higgs Realtime Duplex leads on Big Bench Audio (98.9) but trails GPT-Live-1 on Full-Duplex-Bench (75.5 vs 80.1) and on τ-Voice task completion (69.8% vs 74.5%). Gemini Live is close on some spoken-QA measures but behind on task completion. The key differences are not just scores: Higgs Realtime Duplex is a 45B model with a face and continuous video, GPT Realtime is a shipping product with the widest ecosystem, and Gemini Live is distributed on Android. No independent evaluation has confirmed Boson’s table yet.

Can Higgs Realtime Duplex clone my voice and face?

Yes — that is a designed feature. A short reference audio clip sets the voice and a single reference image sets the face, with no fine-tuning, loaded zero-shot at deployment time. Boson provides default voices and faces, and swapping the references changes the voice and face instantly. Boson has not published consent-check or watermarking details for cloned voices and faces in the Duplex announcement, so check the safeguards before deploying this feature.

Who should care about Higgs Realtime Duplex?

Developers building voice agents, tutors, meeting assistants, game characters, or drive-thru and customer-service agents should watch it closely — especially if they care about interruption handling, cost per hour, and having a visible face. End users should know it exists but wait for the API: the practical step today is to try Higgs Realtime, GPT Realtime, or Gemini Live, and to join the Duplex waitlist if the full-duplex design fits a product you are planning.

The Bottom Line

Higgs Realtime Duplex is one of the most interesting AI launches of November 2026 — not because it tops every benchmark, but because it is built around a genuinely different idea of what a conversation with an AI should feel like. The 160-millisecond chunking, the front model that thinks and calls tools while it talks, the face that moves in sync with the voice, and the tight token budget all point in the same direction: a model designed to be talked to for hours, not queried in turns. The benchmarks say it is competitive with models six times its size, and the demos — meeting assistant, English tutor, drive-thru, hands-free engineer — show a team thinking hard about real products, not just leaderboards.

But the honest verdict has to include the caveats. It is a research preview on a waitlist, with no price and no launch date. Its benchmarks are company-reported and unconfirmed by independent testing, and it does not lead on the benchmark most directly tied to its core claim. The demos are staged, the safety details for voice and face cloning are unpublished, and the lab behind it is small compared with the giants it is competing against. None of that makes Higgs Realtime Duplex a bad bet — it makes it an early one.

If you are a developer, the practical move is clear: join the waitlist, build something small on Higgs Realtime today, and watch how fast Boson ships the Duplex API and at what price. If you are a user, the move is simpler: keep an eye on this space, because the fortnight that brought GPT-Live-1, Soniox STT v6, Phonic-1, and Higgs Realtime Duplex is not a coincidence. Full-duplex voice AI has arrived, and the next year will be decided by which of these models behaves best when a real person interrupts it. Higgs Realtime Duplex is Boson’s answer to that question — and it is a serious one.

Sources: Boson AI, “Higgs Realtime Duplex: Face to Face, Working in the Background” (November 5, 2026), boson.ai; Boson AI Higgs Realtime documentation, docs.boson.ai; Boson AI pricing page; Innovation Essence, “Higgs Realtime Duplex: Boson’s AI That Talks While It Works”; Voice AI Space, “Boson AI announces Higgs Realtime Duplex, a 45B full-duplex omni model”; Piyush Sahoo, “Full-duplex voice models arrived this fortnight” (Medium, November 12, 2026). All benchmark figures are company-reported by Boson AI and have not been independently verified.

Leave a Comment

Your email address will not be published. Required fields are marked *