Voice AI models compared: what the benchmarks say in late 2026

The model behind an AI phone call decides how natural the conversation feels and how often the job gets done. In the four months to October 2026, OpenAI, Google, xAI, Amazon, ElevenLabs and a newcomer, Phonic, all shipped new voice models, and independent benchmarks for them finally exist.

This is a summary for people choosing or buying, not building. Two declarations first. Unibound's AI calls currently run on OpenAI's GPT-Live, so we have a stake, and we have reported the figures as published whether or not they flatter that choice. And every number below was read from its source on 12 November 2026; live leaderboards move, so check the date on anything you quote.

The models at a glance

One design trend is worth knowing. Several of the newest models split the work in two: a fast speech model holds the conversation while a larger model reasons and uses tools in the background. GPT-Live-1, Gemini 3.8 Live's Extended Thinking mode and Phonic all work this way. It is an answer to the old trade-off between sounding quick and being right.

The overall ranking

The most neutral scoreboard is Artificial Analysis's Speech-to-Speech Index, an equal blend of four things: reasoning about what was heard, completing customer service tasks by voice, listener preference in blind comparisons, and task success.

The index was reweighted after its launch in June 2026, so don't compare these with scores from the summer.

Getting the task done

For a business, the component that matters most is τ-Voice: can the agent complete a realistic service task, such as changing an order or rebooking a flight, by voice over a phone-quality line, using tools correctly?

Two things stand out. First, the pace of improvement: the best score on this test was 52.1% in June 2026 and is 79.9% five months later. When Sierra first published τ-Voice in March, voice agents in realistic conditions managed 26 to 38% against 85% for a text agent on the same tasks. That gap has largely closed at the top.

Second, the same model can score very differently depending on its settings. GPT-Live-1 ranges from 59% to 75% depending on which reasoning model sits behind it, and Gemini 3.8 Live more than doubles with Extended Thinking on. "Which model?" is an incomplete question without "configured how?".

Response time

The fastest of the leading models begin speaking in about 0.6 seconds (Eleven Realtime v1 Max at 0.57, Phonic 1 at 0.61), with Grok Voice 3.0 at 0.81 and the OpenAI and Google models between 1.2 and 1.35.

For comparison, people leave about 0.2 seconds between turns, and pauses begin to feel awkward at 0.6 to 0.7. So even the quickest models sit at the edge of natural, before the phone network adds its share. When the testing firm Cekura rang nine vendors' live phone agents, median delays ran from 2.4 to 3.8 seconds. We explain why the gap matters in how AI phone calls work.

Speed and accuracy also pull against each other. The models that think longer, such as GPT-Live-1 on Astra and Gemini with Extended Thinking, complete more tasks and take longer to start talking.

Turn-taking

Knowing when the caller has finished is its own skill, and it is tested separately. On the conversational-dynamics measure Artificial Analysis reports, GPT-Live-1 and Grok Voice 3.0 both score about 97%. Cekura's Turn Bench, which tests end-of-turn detection across six line conditions, puts GPT-Realtime-2.1, Nova 2 Sonic and GPT-Live-1 in a group at 84 to 85 out of 100 with Grok Voice 3.0 at 74, falling to 60 in heavy noise.

The confidence intervals of the top eleven models on that test overlap, so the honest reading is that most leaders are similar in quiet conditions and differ under noise.

Why rankings change with real callers

Benchmarks mostly use simulated callers: another AI playing the customer. Cekura also runs models against recordings of real people, with a stricter pass mark: correct on all three attempts.

Every model drops by more than half, and the order reshuffles. Real people mumble, restart, speak over the agent and have televisions on. Cekura sells testing tools, so it has an interest in showing that testing is needed, but the pattern matches the academic work: in EVA-Bench the typical gap between "passes at least once" and "passes every time" was large.

This is the main reason not to choose on a leaderboard alone, and the main reason a production system needs a person within reach.

What the benchmarks don't tell you

• UK accents. No public benchmark breaks results down by British regional accent. The nearest evidence is that accents in general cost about ten points in τ-Voice, with one model losing eighteen.
• Your tools. Benchmarks use tidy, well-described tools. Yours may be slower and messier.
• Long calls. We found no 2026 study of how quality holds over a ten-minute call.
• Voicemail detection. There is no independent study at production scale, only vendors' own figures.
• Who is measuring. Sierra sells agents, Cekura and Coval sell testing, ServiceNow is a vendor. Artificial Analysis and the university benchmarks are the most neutral.
• How long it lasts. Models are retired quickly. Several realtime models from 2025 are being switched off between November 2026 and January 2027.

How to choose

If you are buying a platform, you are mostly choosing whoever chooses the model for you. The useful questions are:

1. Which model do calls run on today, and how is it configured?
2. What happens when a better one appears? A model change should be a setting. If your prompts, voices and flows have to be rebuilt, you will be stuck on last year's model.
3. Can I hear it on my own phone, with my own script, before I commit?
4. What does a call cost all-in, including the line and the platform? See what an AI voice agent costs in the UK.
5. What happens when the model gets it wrong? No model on any test here is right every time. The answer should involve rules the model can't break and a route to a person.

The models will keep leapfrogging each other; the model that led the main index in June is no longer in its top ten. What lasts is the design around them: a flow that says what happens after each outcome, safeguards enforced outside the model, and your team one step away.

Questions people ask

Sources

1. Artificial Analysis, Speech-to-Speech leaderboard (read 12 November 2026). artificialanalysis.ai
2. Artificial Analysis, announcing the Speech-to-Speech Index (June 2026). artificialanalysis.ai
3. Sierra Research, τ-Voice: benchmarking full-duplex voice agents on real-world domains (March 2026). arxiv.org
4. Cekura, speech-to-speech benchmarks (read 12 November 2026). benchmarks.cekura.ai
5. ServiceNow, EVA-Bench (2026). arxiv.org
6. Full-Duplex-Bench v3 (April 2026). arxiv.org
7. OpenAI, API pricing. developers.openai.com
8. Google, build real-time voice applications with Gemini (September 2026). blog.google
9. xAI, Grok Voice 3 (October 2026). x.ai
10. Phonic, pricing. phonic.ai
11. ElevenLabs, API pricing. elevenlabs.io
12. Coval, speech-to-speech vs cascaded voice AI (November 2026). coval.ai