November 20, 2026
Choose a speech-to-speech model based on its total cost of ownership, not on its price per minute alone. Once you factor in the cost of failed calls, the model with the lowest price per minute can be one of the most expensive per call.
Speech-to-speech models have arrived, and they now serve more production traffic than ever. For most of the past year, no speech-to-speech model surpassed a cascaded pipeline, but that's changed in the past few months.
¹ Pass³ is the share of Cekura's 82 booking and Medicare call scenarios that a model passed on all three of its runs. Cekura measured each score when it tested the model, not on its release day.
Source: Cekura's speech-to-speech leaderboard (November 20, 2026); release dates from each vendor's announcements, docs, and changelogs
There's been a lot of work to assess the quality of these models: benchmarking providers measure naturalness, task success, latency, and more. That said, we think pricing is a nuanced topic and relatively under-discussed. In our experience, buyers often start by comparing per-minute sticker prices across vendors. That's a reasonable first step, but those numbers can hide a lot. The same model can cost very different amounts depending on how it's billed and who sells it, the sticker price sometimes leaves out the intelligence, and not every vendor counts a minute the same way. We'll walk through how to interpret sticker prices and why we believe buyers should think about the total cost of ownership rather than just the cost of the model itself.
You'll typically run into three broad pricing structures for speech-to-speech models:
Source: Vendor pricing pages and docs (November 20, 2026)
Per-token pricing is the most complicated — these models bill the entire conversation on every turn, along with the prompt and tool definitions. Caching is the standard fix (e.g., on GPT Realtime 2.1, the re-read is billed at a small fraction of the price, so the cost per minute stays close to flat throughout the call), but not every model offers this (e.g., Gemini 3.8 Live).
You might work directly with the per-token price if you buy directly from the model vendor, but you may also work with per-minute prices if you buy these models through other vendors. You may find up to a ~6x difference between the cheapest and most expensive per-minute costs (measured, estimated, or listed) for these models. Longer calls, more reasoning, and a platform's own pricing push the cost up, while caching discounts bring it down.
Source: LiveKit, Cekura, Artificial Analysis, and Retell (November 20, 2026)
Vendors often quote a price per minute that doesn't include intelligence, the LLM that runs in the background and primarily handles tool calling. Developers who have built more complex applications on speech-to-speech models will understand that having this intelligence is critical to achieving a sufficient level of performance. For example, GPT-Live-1 advertises 5¢ per minute, but customers can find themselves spending more than 10¢ per minute when they add intelligence.
Tools and telephony are not shown for any vendor.
Source: OpenAI, xAI, ElevenLabs, Retell, and Phonic pricing pages; Cekura; Artificial Analysis (November 20, 2026)
Models like Phonic-1 and Grok Voice 3.0 have the simplest pricing structure: one flat price per minute that includes intelligence for tool calling. That said, with all models, even these models, vendors don't all count a minute the same way. For example, GPT-Live-1 bills per second, without rounding up to a whole minute, and counts the time when "both are silent, or the backend is working," and Phonic-1 takes 95% off silences longer than 10 seconds.
We often find that these nuances don't make an enormous difference at the end of the day, but they can add up. A common example is an agent waiting in silence while navigating an IVR — a vendor that bills silence at the full rate will charge for that time.
Our advice is to look past vendors' sticker prices and start with a trusted third-party benchmark like Artificial Analysis's cost benchmark for speech-to-speech models.
This benchmark runs the same 40 spoken questions on every model and adds up what each one bills — audio and text in both directions, plus reasoning — then reports the cost per hour of the questions' audio.
It isn't perfect. Because it prices single questions rather than whole calls, it leaves out the conversation history that per-token models bill again on every turn, as well as caching discounts and tool costs. It also occasionally misses the cost of reasoning.
Still, we believe it's the most helpful starting point before you test models on your own calls.
Source: Artificial Analysis, cost per hour of input audio, converted to cents per minute (November 20, 2026)
Taking a step back, we'd argue that cost per minute alone isn't sufficient to choose a model to build on. Instead, you should compare the total cost of ownership, which takes into account how often the model succeeds for your use case.
Let's take a hypothetical example of a company building a voice agent to field inbound customer support calls:
For this company, the cost of a call is the cost of the four minutes plus the cost of a person resolving the calls the agent couldn't. So to compare models, you need to know how often each one fails.
You should run evals on several models to see how well they complete calls for your use case, but for this hypothetical, we can use Sierra's τ-Voice 2.0 benchmark as a rough proxy. This benchmark runs models through 150 simulated support calls (retail orders, airline reservations, telecom support), and a task passes only if the account ends in exactly the right state (some tasks also check the agent's actions and what it told the caller). There's a lot that goes into making a voice agent successful, but this benchmark pulls out the differences that come from the model, its reasoning setting, and sometimes the reasoning model behind it. We find that there's a wide variance: success rates run from 36% to 86%.
A task passes only if the booking, order, or account ends up exactly right; some tasks also check what the agent did or told the caller.
Source: Artificial Analysis, τ-Voice 2.0 task success (November 20, 2026)
If you look at this data in tandem with the price-per-minute data, you can approximate the total cost of ownership: the cost of the minutes plus the cost of having a person resolve the calls the model fails.
This company might be debating two different configurations of GPT-Live-1: Astra (medium) and Sol (low). Astra (medium) costs meaningfully more at 9.72¢ per minute vs. 7.45¢ per minute (per Artificial Analysis), but it successfully handles 80.4% of calls vs. 74.7%. Which is the right answer? The true cost of each is the cost of its minutes (price per minute × 4 minutes per call) + the cost of its failed calls (chance it fails to resolve the issue × $13.50 for a person to handle it).
Source: τ-Voice 2.0 results and cost per hour of input audio from Artificial Analysis (November 20, 2026); Gartner's 2024 cost-per-contact benchmark
For each customer support request, Astra (medium) costs 9¢ more in minutes but 77¢ less in failed calls, so the company would be better off going with Astra (medium) in this case.
$13.50 is Gartner's 2024 median cost of an assisted contact (phone, chat, or email), used as the cost of a failure.
Source: τ-Voice 2.0 results and cost per hour of input audio from Artificial Analysis (November 20, 2026); Gartner's 2024 cost-per-contact benchmark
The chart above runs the same math for all eight models, and the ranking looks very different from the ranking by cost per minute. The cheapest model per minute makes for one of the most expensive systems in this scenario: Gemini 3.8 Live costs 1.4¢ per minute but $8.75 for a full support interaction, more than every other model except GPT Realtime 2.1 (high).
You could additionally calculate the break-even: the cost of a failed call that makes it worthwhile to choose one model over another. You could compare two extremes: Phonic-1 at 7¢ per minute and an 85.6% task success rate vs. Gemini 3.8 Live at 1.4¢ per minute and a 35.6% task success rate. If a failure costs more than 45¢, it's worthwhile to go with Phonic-1, and vice versa.
Source: τ-Voice 2.0 results and cost per hour of input audio from Artificial Analysis (November 20, 2026)
Not every call is a customer support call, but most calls have some quantifiable costs associated with a failure. For example, if your voice AI books jobs for a local home services business, a failed call can cost you the job, which goes to another provider, plus the cost of an overflow call center. If it onboards new users for a consumer app, a failed call costs roughly what you paid to acquire that user.
In theory, this math should also include costs that are harder to quantify. A model with slow responses or an unnatural voice can successfully complete a call but leave a customer unhappy, which shows up as lower NPS and higher churn. In practice, those costs are hard to assign a number to, so we keep them out of the math and use latency and naturalness to choose among the models that cost the least.
If you're in the market for a speech-to-speech model, here's what we recommend:
Of the eight models in the example above, Phonic-1 has the highest success rate and the lowest total cost. In production, CloudTrucks used Phonic to raise containment, the share of calls its agents finish without a person, from 31% to 85% in a month, and saved $2.2M a year in human support costs. To see what Phonic can do for you, try Phonic on the calls your agent struggles with today. Sign up by November 30, 2026 for $250 in free credits.