All insights
The Market6 min read

The Voice Is the Cheap Part

Vishal Sachar

Co-Founder & CEO of CLRT

On 10 September, OpenAI put a voice that talks and listens at the same time into its API at five cents a minute, billed by the second. For a sales or customer-experience leader, the arithmetic now looks settled. An hour of natural conversation costs three dollars in voice, it never sleeps, and the board has been asking for an AI plan with a number attached. The obvious move is to buy a voice agent and point it at the phone. That reading is understandable and mostly wrong. A price this low does not mean voice agents have become easy. It means the answer itself has stopped being worth anything, and whatever made one conversation sell while another stalled has moved somewhere the meter does not reach.

$0.051
Per minute for OpenAI's full-duplex GPT-Live-1 voice in its API, billed per second; backend model and tools extra
OpenAI, 2026
32%2
Banking voice support pass rate OpenAI reports for its newest voice model, up from 12.4% (vendor figures)
The Decoder, 2026
40%3
Share of the average seller's time spent actually selling, in Salesforce's State of Sales survey of 4,050 sales professionals in 22 countries (vendor research)
Salesforce, 2026

The cheap minute did not arrive alone. On 13 October 2025 Salesforce announced Agentforce Voice, a native voice layer meant to turn the old phone menu into a real-time conversation, wired into the Salesforce platform. On 4 February 2026 ElevenLabs raised 500 million dollars at an 11 billion dollar valuation, reporting, by its own account, more than 330 million dollars of annual recurring revenue at the end of 2025 and naming its agents platform among its priorities. In September OpenAI made the voice a metered utility. When a CRM incumbent, a voice specialist and a model lab all sell the same capability within eleven months, and one of them prices it by the second, the capability has stopped being an advantage. It has become infrastructure, like the phone line itself. Nobody ever won a sale because their line was clear.

FIG. 01Three moves in eleven months turned the voice from a feature into a utility. Sources: Salesforce, 13 Oct 2025; ElevenLabs (vendor), 4 Feb 2026; OpenAI, 10 Sep 2026.
01What the meter counts

Read the price page closely and it says more than the headline. Voice sessions cost five cents a minute; the backend model and the tools it calls are billed separately. That sentence is the whole market in miniature. The meter counts the hearing, the speaking and the timing, the part that makes a conversation feel natural. It does not count the part that decides what to say, looks up the customer, checks the policy or does anything once the talking stops. Those are priced elsewhere, and more importantly they are designed elsewhere, by whoever deploys the agent. A buyer comparing voice agents on price per minute is comparing the one component every competitor can now buy at the same rate, and ignoring the components that decide whether anything happens on the call at all.

The vendor's own scorecard makes the same point more sharply. OpenAI reports that its new model lifts full-duplex interactivity, a benchmark of speaking and listening at once, from 45.4 to 80.1 percent against its predecessor, cuts turn-taking latency from 1.4 seconds to 0.8, and, with a reasoning model behind it, raises tool-calling accuracy from 60 to 87 percent. These are real gains in how a conversation feels, and they are the vendor's own launch figures. Read the one number closest to the job honestly: on a banking knowledge test its maker chose, the newest voice passes 32 percent of tasks, up from 12.4, roughly one in three. The vendor's stronger task scores come from setups that pair the voice with a separate reasoning model doing the thinking. That is the point. The talking has become good, but the job was never in the voice. It lives in knowing the customer, the product and the rule, and in what the agent is permitted to do next.

FIG. 02OpenAI's own launch figures: the conversation got far better, the banking job much less so. Vendor benchmarks, reported by The Decoder, 10 Sep 2026.
02The scarce hour

Salesforce's latest State of Sales survey, vendor research drawn from 4,050 sales professionals in 22 countries, finds the average seller spends 40 percent of their time selling. Fifty-four percent have already used AI agents, and nearly nine in ten plan to by 2027. The usual pitch reads this as a case for automating the other 60 percent. The sharper reading is that the 40 percent is the scarce asset: the live conversation in which interest becomes a decision. A cheap voice pointed at answering buys back very little of that hour, because answering was never where the hour went. Pointed at the conversation itself, at understanding what this buyer needs and moving them to a next step, an agent could extend the one part of the week that actually produces revenue.

FIG. 03The average seller spends 40% of their time selling. Salesforce State of Sales 2026, vendor survey of 4,050 sales professionals, self-reported.
03Where value moved

So the value has moved inside the call, into four things no meter prices. Understanding: hearing what the buyer asked and what they did not say. Showing: putting the thing itself in front of them while it is being discussed, because a voice alone asks a buyer to imagine, and imagining is where doubt grows. Meeting the doubt: answering the objection at the moment it is spoken, with the proof to hand, rather than promising a callback. Taking the next step: the viewing booked, the proposal sent, the payment taken before the conversation ends. Each of these is a decision about one specific business: what may be promised, what must be shown, when a human takes over. That is where a voice agent is won or lost, and it is the part no vendor can sell you by the minute.

The meter counts the hearing and the speaking. It does not count the part that sells.

A deeper dive

The mechanism is a quiet inversion in how voice agents get bought. When a conversational voice was expensive, the business case was cost per call: replace an expensive human minute with a cheaper machine minute, and the agent that absorbed the most minutes looked like the best investment. At five cents a minute that case shrinks to a rounding error, and the measure that remains is what each conversation produced. A call that ends with a booked viewing is worth something; a pleasant call that ends with a promise to send some information is worth nothing at any price. Yet most procurement still compares voice agents on the old axis of latency, naturalness and price per minute, the very properties that have just converged across vendors. The buyer ends up choosing between near-identical voices on the dimensions that no longer differ, while the dimensions that do differ, what the agent knows, what it can show, which commitments it may make and how it hands over to a person, appear nowhere in the evaluation because no datasheet lists them. Those are decisions about your business, and only you, or whoever designs with you, can make them.

The second-order trap is volume. Cheap minutes invite a firm to point the voice at everything: every inbound enquiry, every dormant contact, every after-hours call. Fluency then becomes a liability, because a fluent agent that cannot do the job loses a buyer faster, and at greater scale, than a slow human who can. Reliability compounds the problem. In tau-bench, a vendor benchmark Sierra published in 2024, the strongest function-calling agent of the day completed about six in ten retail tasks and about a third of airline tasks on a first attempt, and in retail the share it got right on all eight repeated attempts fell below a quarter. A demo is one good call. Production is the thousandth, with an interruption, a change of mind and a question nobody scripted. Cheap voice makes it easy to reach the thousandth call quickly, which is exactly when the gap between a voice that answers and an agent that sells shows up in the numbers, usually after the contract is signed. The firms that win this cycle will spend what they save on minutes on the design inside the conversation, not on more minutes.

Work with CLRT

CLRT built Closer for the part after the answer. Closer, our immersive sales agent, puts what the buyer is asking about on their screen while it speaks, and takes the next step before the conversation ends. Before you buy a voice agent because it suddenly looks cheap, talk to Tom, the agent itself, at clrtstudio.com/closer. Then talk to CLRT about which of your conversations deserves an agent, and what must happen inside it, because that design, not the voice, is what decides whether the call sells.

Vishal Sachar

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.

Start here

Skip the reading. See where your leverage leaks.

Ascent is our free diagnostic. Ten minutes, and you have the one workflow worth building first.