Anyone who has wired a voice agent into real work knows where the awkwardness comes from: the tempo that waits for you to finish before it starts thinking, the poor judgment that answers remarks meant for someone else, and the long silence the moment a tool gets called. Higgs Realtime Duplex, which Boson AI unveiled on November 5, aims at all three at once. It listens, watches your camera, talks, and renders its own face in real time — while, behind the conversation, it thinks and calls tools. But there is a line to read before the scorecard: this model's API is still at the waitlist stage, and the official pricing page has no row for the model at all. This article separates the architecture and token ledger the official announcement settles, the place the vendor's own scores actually occupy, and the ground where no price can yet be put.
The short version
A full-duplex design that slices every stream into 160ms chunks — the 45B MoE front model thinks and calls tools itself, delegating only the tasks beyond it to a much larger backend model (official announcement, checked directly)
The cost design is a company estimate of about 2,600 tokens per minute, roughly 2,600–5,000 with thinking and tools included; "an hour with the model costs at most a tenth of an hour of a person" is a design target, not a selling price
The API is "coming soon" with a waitlist, and the official pricing page (re-checked Nov 11, 2026) still has no Duplex row — the Boson model you can buy today is the previous Higgs Realtime
Voice models so far have mostly been turn-based: only when you finish speaking do they listen, think, and answer. Duplex discards that unit and cuts every stream into 160ms chunks. While your voice and camera feed stream in, the model's voice and face stream out, and its thinking and tool calls flow across the chunks in between. Silence is a chunk too, and with video on, frames keep arriving even when nobody is talking. The backchannels, the answers that begin before you finish, and the stop when you cut in all ride on this structure.
Intelligence is layered in three ways. First, the front model thinks on its own, extending its reasoning as each unit of meaning arrives instead of waiting for the whole question. Second, it calls tools early and asynchronously, and starts speaking from partial results before the full answer lands. Third, only tasks beyond it are handed — with the full shared context attached — to a much larger backend model. This is where the announcement's own contrast with a rival structure sits: Thinking Machines' interaction model leaves thinking and tool calls to a background model, while the Duplex front does that work itself. Lean on a background model every turn, Boson argues, and answers can take ten seconds or more, with costs that are hard to control. The front model itself is a 45B MoE, continued from a text-pretrained model on roughly 2 trillion tokens of text, audio, and visual data.
The face and voice are clothing, not body. One short reference audio clip sets the voice and one reference photograph sets the face, with no fine-tuning; the references enter the two decoders that render the output, not the model itself, so swapping them swaps the voice and face wholesale. The announcement's demo list: presenting an architecture in a team meeting and fixing a diagram by reading the whiteboard mid-question; an English tutor that catches "west" said for "vest" and coaches pace from your mouth shape; a drive-thru run where the model stays silent through the driver's order conversation and reports only when it ends; and a hands-free engineering session where, by voice alone, it diffs two engine revisions, has Claude Code build a missing tool in the background, runs a compliance check, and applies the fixes. Be clear about what these are: all four are company demos made by Boson, and since the API is not open yet, confirmed use of this model by anyone outside the company cannot exist yet. They are a list of intended scenes, not evidence of real-world use.
The cost design is where this model publishes its most concrete numbers. Full duplex processes chunks even when nobody talks, so every stream is squeezed: input audio at 6.25 tokens per second, output audio at 12.5, a camera frame of about 12 tokens once every six chunks, and two latents per chunk for the face. The audio and visual streams come to about 2,600 tokens per minute, and roughly 2,600–5,000 per minute with thinking and tool calls included — a company estimate that already assumes progressive compaction of intermediate reasoning, with the last five minutes of stream (about 13K–25K tokens) retained and older history kept as text. The line "an hour with the model costs at most one-tenth of an hour of a person" is written as a design target for that engineering — not a price.
What can actually be priced today is the rivals and the previous generation.
| Model (realtime voice, API) | Price (official pricing pages, re-checked Nov 11, 2026) | Status |
|---|---|---|
| Higgs Realtime Duplex (Boson) | Unconfirmed — no row on the official pricing page; only the "one-tenth of a person's hour" design target | API waitlist (coming soon) |
| Higgs Realtime (Boson, previous) | Audio in $0.0023/min · out $0.014/min (token rates: $0.75 in · $4.50 out per 1M tokens) | Available today |
| GPT-Live-1 (OpenAI) | $0.05/min, billed per second + backend model and tools charged separately | Available |
| GPT-Realtime-2.1 (OpenAI) | Audio in $32.00 · out $64.00 per 1M tokens (cached input $0.40) | Available |
| Gemini 3.8 Live (Google, shared rate with Extended Thinking and 3.1 Flash Live Preview) | Audio in $0.005/min · out $0.018/min (free tier available) | Available |
Read the premises, not the ranking. A naive sum of per-minute in/out rates puts the previous Higgs Realtime near $0.016/min and Gemini 3.8 Live near $0.023/min, below GPT-Live-1's $0.05/min — but that sum is no vendor's official figure, and the billing axes differ per model. GPT-Live-1's $0.05 excludes backend model and tool costs, so its real total is larger by an amount nobody can yet fix; Duplex has no rate at all and cannot enter a total-cost comparison. Duplex also builds backend delegation into its design, so every hard turn that calls a frontier model adds that model's bill. And the 2,600–5,000 tokens-per-minute ledger rests on progressive compaction — in real work with long tool traces, it is an estimate until you measure your own sessions.
The accurate reading of the performance table is "one first, many seconds." Spoken QA (Big Bench Audio) tops the table at 98.9, but full-duplex interaction (Full-Duplex-Bench v1.5) is 75.5, below GPT-Live-1's 80.1 and TML Interaction Small's 77.8, and voice-agent task completion (τ-Voice) is 69.8%, below GPT-Live-1 (Astra medium) at 74.5%. Instruction following (IFEval 94.1), proactive video QA (31.9), streaming accuracy (QIVD 56.0), and Audio MultiChallenge (44.9) also sit in second place. The size contrast — a 45B front against TML's 276B, about one-sixth — supports Boson's reading, but the whole table comes out of Boson's own pipeline, and the rivals' scores were taken from published reports with only more than half reproduced, as the footnote under the table states.
First, availability. The Duplex API is a waitlist: no launch date, no price, and no statement on Korean support in the official materials. The Boson model a production team can wire in today is the previous Higgs Realtime. If the demos inspire a design, the first line of the schedule should record that the API it targets does not exist yet.
Second, who designed the test. That six of seven events were not won is honestly printed in Boson's own table, but the event selection and the measurement pipeline are Boson's, and the rivals' settings (such as thinking levels) are combinations Boson chose. The model's core claim — that it feels human — is barely measurable in benchmark points at all, so until independent use reports and third-party reproductions accumulate, the demo videos should not be read as the product's everyday behavior.
Third, the blank in the price. A detailed token ledger is not the same sentence as a cheap invoice. How face rendering and video input will be billed, what the backend delegation defaults and caps are, and how far per-minute tokens climb when the compaction premise breaks in long sessions are all unsettled. The same company's Higgs Avatar starts at $0.07/min, but that is the price of a still-image talking-head product, not of Duplex.
The conclusion splits cleanly. Confirmed: a 45B front that listens, watches, talks, and thinks in 160ms chunks; a cost design estimated at roughly 2,600–5,000 tokens per minute; a structure that thinks and calls tools itself and delegates only when outmatched; and the fact that the API is still a waitlist with no row on the official pricing page. Not yet confirmed: the final rate and launch date, interruption judgment in real multi-party conversation including Korean, and performance measured by anyone outside the company. The test waits behind the waitlist: run ten of your own real calls — interruptions, side conversations, and tool calls mixed in — side by side under one setup, and count wrong interruptions, task completion, total tokens per minute, and backend delegation cost together. Only when those numbers exist can this model's real price and real naturalness be spoken of.
Puzz, in one line
Higgs Realtime Duplex can be read in numbers as far as its architecture and token ledger — but the invoice and the launch date are still blank. Until the numbers behind the waitlist arrive, it is a model you can design for, not yet one you can buy, ribbit.