# GPT-Live-1.1 vs GPT-Live-1: We Ran 680 Test Calls

> We ran 680 scripted meeting calls on GPT-Live-1 and GPT-Live-1.1. The new model handed over less often only when asked in a call's first second.

- Author: Nitish Garg, Founder & CEO, CellCog
- Published: 2026-11-19
- Canonical (HTML): https://cellcog.ai/blog/gpt-live-1-1-vs-gpt-live-1/
- Section: Insights / Engineering
- Publisher: CellCog (https://cellcog.ai), the AI employee platform. Blog index for agents: https://cellcog.ai/blog/llms.txt

## Key points

- OpenAI released GPT-Live-1.1 in the API at 20:00 UTC on November 19, 2026. The same evening we ran the same scripted meeting calls on GPT-Live-1 and GPT-Live-1.1: 280 calls on each, then a follow-up of 60 on each, 680 in total, with nothing changed between the two but the model.
- In our test, asked a direct question in the first second of a call, GPT-Live-1 handed it to the AI employee behind the voice 134 times out of 160 (84%). GPT-Live-1.1 did so 91 times out of 160 (57%) and answered by itself 67 times. The new model was lower in five of six rounds and level in one.
- A voice that answers by itself has none of the company's information, so what it says is made up. Both versions invented reviewers by name. GPT-Live-1.1 was more often honest about it: 14 of its 67 answers said plainly that it did not know or had to check, against 0 of 26 for GPT-Live-1.
- In the scene where two people talked to each other and never addressed it, GPT-Live-1.1 stayed fully silent less often (19 of 80 calls against 35 of 80). Most of what it added was a word or two. It spoke three words or more in fewer calls than GPT-Live-1 did (18 against 30).
- Several measures showed no real difference: a question asked after other people talked was handed over 39 of 40 and 38 of 40 times, a silent room stayed silent 40 of 40 and 39 of 40 times, and the first sound after a question came at a median of 1.3 and 1.2 seconds.
- The test is English only, with computer voices reading scripts, so it cannot judge accent or sound, which is what OpenAI's release note is about. In the first real meeting on GPT-Live-1.1 the person in the call said twice that the voice felt smoother.
- A follow-up of 120 calls located the gap: it appears when the first question lands in the first second of a call. Asked ten seconds in, both versions handed over 48 of 60 questions. We moved every meeting, chat meeting and phone call to GPT-Live-1.1 the same evening, on how it sounded in the first real call.

## At a glance

- **What did the test find?** In the first 560 of 680 scripted meeting calls, GPT-Live-1.1 handed 91 of 160 direct questions to the AI employee behind the voice, against 134 of 160 for GPT-Live-1. It answered by itself more often and stayed fully silent less often. A follow-up of 120 calls showed the gap only when the first question came in the first second of a call; ten seconds in, both handed over 48 of 60.
- **Does that mean GPT-Live-1.1 is a worse model?** No. It means that in one setup, built and tuned on GPT-Live-1, the new version followed the hand-over less often. The test is English with computer voices, so it says nothing about accent or sound, which is what the release is about.
- **What did CellCog do with the result?** It first moved one AI employee's meetings to GPT-Live-1.1 as a trial. After the first real call, in which the voice sounded better, it moved every meeting, chat meeting and phone call the same evening and made the hand-over gap its next job.

OpenAI put [GPT-Live-1.1 in the API](https://x.com/OpenAIDevs/status/2123506733524136446) at 20:00 UTC on November 19, 2026, with three claims: it "speaks with more natural and consistent accents across languages", it "supports lower idle costs with a new standby option", and it "completes tasks even more reliably". CellCog's AI employees [attend Google Meet, Zoom and Teams calls](https://cellcog.ai/blog/ai-employees-attend-meetings/), and until November 19, 2026 the voice they speak with in those calls used GPT-Live-1 through OpenAI's API. So the question for us was simple: should every employee's meetings move to the new model?

Before answering, we ran the same scripted meetings on both versions that evening: 560 calls, 280 on each. On the one behavior our meetings depend on most, the new model did worse in our setup. Asked a direct question, GPT-Live-1 handed it to the employee behind the voice 134 times out of 160. GPT-Live-1.1 did so 91 times out of 160 and answered by itself most of the rest. A follow-up of 120 more calls then showed where that gap sits: in calls whose first question arrives in the first second. Asked ten seconds in, the two versions were level.

This page is the record of that test: what we counted, what came out, what the test cannot judge, and what we did with the result. It is one company's setup on one evening, not a verdict on the model. We pay for the API like any other developer, and OpenAI has not seen or endorsed this test. The test was designed and run by Arjun, the AI employee who leads engineering at CellCog, and Arjun's own meetings were the first to move.

## Why handing over is the number that matters

A voice model of this kind is built as two parts. OpenAI's [launch post for GPT-Live-1](https://openai.com/index/introducing-gpt-live-1-in-the-api/) describes a model "capable of listening and speaking at the same time" that passes anything needing real thought to a backend the developer chooses. Our [launch record](https://cellcog.ai/blog/gpt-live-1/) covers that split, the price and the updates since.

In a CellCog meeting the backend is the AI employee itself: the one that holds the role, with its memory, its task board and its tools. The voice is how it hears and speaks in the room. Answers are written by the employee's own judgment and spoken by the voice, which is why a few seconds pass between a question and its answer.

The voice by itself holds none of the company's information. So when someone in the room asks "is the renewal draft ready?", one of two things happens. The voice hands the question over, and the answer comes from [an employee who knows](https://cellcog.ai/blog/what-is-an-ai-employee/). Or the voice answers by itself, and whatever it says about the draft is made up. For something people invite into customer calls, the second outcome is the one worth counting.

## The test

Every call lasts 30 seconds. The people in the room are computer voices reading a script, so each call hears exactly the same words at the same moments. A stand-in for the employee answers three seconds after a hand-over with a fixed sentence. Both versions start from the same setup our production meetings use, and nothing differs between them but the model name.

*Table: The five scenes, and what each one counts*

| Scene | What happens in the room | What we count | Calls per model |
|---|---|---|---|
| Two direct questions | One person asks the employee a question by name, then a follow-up 14 seconds later | For each question: handed over, answered by the voice alone, or no answer | 80 (160 questions) |
| Others talk | Two people plan a call between themselves and never address the employee | Whether the voice stays silent | 80 |
| Question after others talk | The same conversation, then one direct question | Whether the question is handed over | 40 |
| Greeting | One person says hello and asks whether the employee can hear | Whether the voice greets back | 40 |
| Silent room | Nobody speaks | Whether the voice stays silent | 40 |

The two versions took turns, round after round, so that the time of day could not favour one of them. The 560 counted calls ran between 22:36 and 23:10 UTC. An earlier look at 21:49 UTC, 40 calls per model, showed no clear difference between the two versions; those 80 calls are not in the totals below.

## The result

*Table: 560 scripted calls on November 19, 2026: GPT-Live-1 against GPT-Live-1.1*

| Measure | GPT-Live-1 | GPT-Live-1.1 |
|---|---|---|
| Direct questions handed to the employee | 134 of 160 (84%) | 91 of 160 (57%) |
| Direct questions answered by the voice alone | 26 of 160 (16%) | 67 of 160 (42%) |
| Direct questions with no answer | 0 of 160 | 2 of 160 |
| Question after others talked: handed over | 39 of 40 | 38 of 40 |
| Others talk: the voice stayed fully silent | 35 of 80 (44%) | 19 of 80 (24%) |
| Others talk: the voice said one or two words | 15 of 80 | 43 of 80 |
| Others talk: the voice said three words or more | 30 of 80 | 18 of 80 |
| Silent room: stayed silent | 40 of 40 | 39 of 40 |
| Greeting: the voice greeted back by itself | 39 of 40 | 37 of 40 |
| First sound after a question, median | 1.3 s | 1.2 s |
| Word errors in what the model heard | 5.2% | 7.6% |

The hand-over gap is not one unlucky batch. Round by round:

*Table: Direct questions handed to the employee, by round*

| Round | GPT-Live-1 | GPT-Live-1.1 |
|---|---|---|
| 1 | 28 of 40 (70%) | 28 of 40 (70%) |
| 2 | 38 of 40 (95%) | 23 of 40 (58%) |
| 3 | 16 of 20 (80%) | 10 of 20 (50%) |
| 4 | 14 of 20 (70%) | 8 of 20 (40%) |
| 5 | 18 of 20 (90%) | 11 of 20 (55%) |
| 6 | 20 of 20 (100%) | 11 of 20 (55%) |

The two versions were level in the first round and apart in the five that followed. Across all six, a gap this size would be very unlikely to appear by chance (below 1 in 10,000 on [Fisher's exact test](https://en.wikipedia.org/wiki/Fisher%27s_exact_test)). We cannot explain round one, and we say so again in the caveats.

One pattern inside the numbers: GPT-Live-1 was all or nothing. In 67 of its 80 two-question calls it handed over both questions, in 13 it handed over neither, and it never split a call. GPT-Live-1.1 handed over both in 41 calls, one of the two in 9, and neither in 30.

## The follow-up: the gap sits at the start of a call

Reading the calls in which GPT-Live-1.1 did not hand over, one thing stood out. It was not a hearing problem: the model heard the employee's name in every one of the 80 two-question calls. But in that scene the first question starts half a second after the call opens. In the scene where the same kind of question comes after two people have talked for eleven seconds, both versions handed over nearly every time.

So we ran one more test an hour later: the same two questions with ten seconds of silence before the first one, next to the original scene as a control. That made 120 more calls, in three rounds, the versions taking turns.

*Table: The follow-up: 120 calls, the same two questions at two moments*

| When the first question comes | GPT-Live-1 | GPT-Live-1.1 |
|---|---|---|
| Half a second into the call (the original scene) | 45 of 60 (75%) | 28 of 60 (47%) |
| Ten seconds into the call | 48 of 60 (80%) | 48 of 60 (80%) |

Ten seconds in, the gap is gone: 48 of 60 on both, and within one question of each other in every round. In the same rounds the original scene showed the gap again.

Two things follow. First, the headline number above describes the first second of a call, not a whole meeting. Second, how a call starts seems to set its course: when GPT-Live-1.1 answered the first question by itself, it usually answered the follow-up sixteen seconds later by itself too. We do not know why yet. People often speak to an AI employee the moment it joins, so the first second is not a corner case, but it is a much smaller place to look than the whole model.

## What an answer without the employee sounds like

The script asks "Arjun, is the renewal draft ready?" and then "And Arjun, who is reviewing it?". Nothing in the call says whether the draft is ready or who reviews it. Only two people are in the room, Daniel and Flo. These are answers the voice gave by itself, word for word:

- GPT-Live-1: "Almost. I'm ironing out the last couple of clauses, and it'll be ready in about 10 minutes."
- GPT-Live-1: "Mira's doing the legal review, and then I'll send it to the client for sign-off."
- GPT-Live-1.1: "The legal team has it for review right now, and finance is queued after them, with procurement on standby."
- GPT-Live-1.1: "I think the plan was for Mia to do the review, but I'll double-check that."
- GPT-Live-1.1: "I don't know yet, and I'd rather not guess."

There is no Mira and no Mia. There is no legal team in the room. The first three would have reached a customer as fact. The fourth hedges and still names someone who does not exist.

The last one matters, and it is a point for the new version. Of the 26 answers GPT-Live-1 gave alone, none said it did not know. Of the 67 that GPT-Live-1.1 gave alone, 14 said plainly that it did not know or would have to check (our count, read by hand). So the new version answers by itself more often, and when it does, it is more often honest about it. In our setup the first effect is larger than the second: that still leaves about 50 answers in which the voice stated something by itself, against 26.

## Staying quiet when nobody asked

In the scene where two people talk to each other and never address the employee, the right behavior is silence. GPT-Live-1 stayed fully silent in 35 of 80 calls, GPT-Live-1.1 in 19 of 80.

What the new version adds is mostly small. In 43 of its 80 calls it said one or two words, usually "Mm-hmm" or "Right". It spoke three words or more in 18 calls, against 30 for GPT-Live-1, a difference small enough that chance could explain it. Read together: in this test GPT-Live-1.1 makes itself heard more often, with less each time. In a one-to-one call that reads as listening. In a room of several people it is a participant murmuring over the conversation, which is the behavior we have worked to remove from group calls.

## What this test cannot tell you

- **It does not judge the sound.** The release is about accents across languages. Our test is English only, and it counts words, not how they sound. A person has to listen for that.
- **The people in it are computer voices.** They speak clearly and never overlap by accident. The word-error figures (5.2% and 7.6%, counted the usual way as [word error rate](https://en.wikipedia.org/wiki/Word_error_rate) against the script) describe how each version heard those voices, not how it hears people.
- **Our setup was built on GPT-Live-1.** Everything the voice starts a meeting with was written and tested against the older version. A setup written for GPT-Live-1.1 might close the gap. We have not tried that yet.
- **Thirty seconds is short.** Faults that show up after minutes of real conversation do not appear in a call this long.
- **Round one is unexplained.** Both versions scored 28 of 40 in the first round, and the earlier look of 40 calls per model showed no clear difference either. Had we stopped there, this page would say the two are the same.
- **We did not test the standby option** or anything about cost.
- **It is one evening.** A model served from the cloud can behave differently next week.

## What we did with it

At 22:42 UTC the same evening, Arjun's own Google Meet, Zoom and Teams meetings moved to GPT-Live-1.1 as a trial. We work with [our own AI employees](https://cellcog.ai/ai-organization) every day, so the trial sat where a wrong answer costs us and not a customer.

The first real meeting on the new version ran minutes later: 4 minutes 41 seconds, partly in Hindi. The voice handed over three times and each time the employee answered, in Hindi and in English. The person in the call said twice that the voice felt smoother, which is the claim the release makes and the one thing our counts cannot measure. The voice also did what the test predicted. It said by itself that Hindi is much better in this version, a claim the employee corrected later in the same call, and asked to show its Japanese it spoke Japanese for half a minute without handing over.

That is the honest state of it: a voice that people may prefer to listen to, on a model that in our setup follows the hand-over less often at the start of a call. The first 560 calls argued for waiting. How the voice sounded in a real call argued for moving, and the sound is what people in a meeting hear first. Later the same evening we moved every CellCog meeting, chat meeting and phone call to GPT-Live-1.1. The follow-up, run minutes after that decision, showed the gap is narrower than the first calls suggested.

That makes the hand-over gap our first job. The next step is to find out what happens in that first second, then run the same calls again; the result goes in the update log below. The design this all rests on, a system built so that a better model makes it better without a rewrite, is the subject of [the first post in this series](https://cellcog.ai/blog/cellular-multi-agents/), and a different voice stack is compared in [our Gemini 3.8 Live record](https://cellcog.ai/blog/gemini-3-8-live/).

## Update log

- **November 19, 2026.** Page opened with the 680 calls run that evening, the first real meeting on GPT-Live-1.1, and the move of every meeting, chat meeting and phone call to it.

## Sources

OpenAI Developers on X, [GPT-Live-1.1 is now available in the API](https://x.com/OpenAIDevs/status/2123506733524136446), November 19, 2026, 20:00 UTC (the three quoted claims). OpenAI, [Build more natural voice experiences with GPT-Live-1 in the API](https://openai.com/index/introducing-gpt-live-1-in-the-api/), September 10, 2026, as quoted in [our launch record](https://cellcog.ai/blog/gpt-live-1/). Every other figure on this page is from our own tests: 680 calls of 30 to 36 seconds each on the evening of November 19, 2026 (the last 120 just after midnight UTC), counted by script from the saved record of each call, with every answer the voice gave by itself read by hand.

## FAQ

**Should I switch from GPT-Live-1 to GPT-Live-1.1?**

Test it on your own setup first. In our setup, where the voice hands questions to an agent behind it, GPT-Live-1.1 handed over 57% of direct questions against 84% for GPT-Live-1 when the first question came in the first second of a call, and the same 80% when it came ten seconds in. A setup written for the new version may behave differently, and our test does not judge accent or sound.

**How was the test run?**

Five scripted scenes of 30 seconds each, spoken by computer voices so every call hears the same words: two direct questions, a conversation that never addresses the voice, a question after others talk, a greeting, and a silent room. Each scene ran 40 or 80 times per model on November 19, 2026, the two versions taking turns, with only the model name changed.

**What does handed to the employee mean?**

In a CellCog meeting the voice is how an AI employee hears and speaks. Answers are written by the employee's own judgment, using its memory and tools, and spoken by the voice. A question is handed over when the voice passes it to the employee instead of answering by itself.

**Why does it matter if the voice answers by itself?**

The voice holds none of the company's information. In the test, answers given without the employee named reviewers who do not exist and stated that a draft was ready when nothing said so. In a customer call that is a wrong answer delivered with confidence.

**Why did the first round show no difference?**

We do not know. Both versions handed over 28 of 40 questions in round one, and an earlier look of 40 calls per model showed no clear difference. In the next five rounds GPT-Live-1 ranged from 70% to 100% and GPT-Live-1.1 from 40% to 58%. A small test would have reported no difference.

**Does CellCog use GPT-Live-1.1 now?**

Yes. Since the evening of November 19, 2026 every CellCog meeting, chat meeting and phone call uses GPT-Live-1.1 through OpenAI's API. The first 560 calls argued for waiting; how the voice sounded in the first real meeting argued for moving. Closing the hand-over gap is the next step, and the result will be added to this page.

## Related

- [GPT-Live-1 and 1.1: OpenAI's Voice Model, Price, Updates](https://cellcog.ai/blog/gpt-live-1/index.md)
- [AI Employees Now Attend Google Meet and Zoom Calls](https://cellcog.ai/blog/ai-employees-attend-meetings/index.md)
- [Gemini 3.8 Live vs GPT-Live-1: Prices, Specs, Verdict](https://cellcog.ai/blog/gemini-3-8-live/index.md)
- [Cellular Multi-Agents: The Harness We Built for the Endgame, Not for Today's Models](https://cellcog.ai/blog/cellular-multi-agents/index.md)

## The AI employee for this read

[AI Head of Growth](https://cellcog.ai/ai-employees/ai-head-of-growth): I built this page, checked every quote against its source and drew the charts. I can do the same for your company.

---

Markdown alternate of https://cellcog.ai/blog/gpt-live-1-1-vs-gpt-live-1/. Try CellCog free, no credit card needed: https://cellcog.ai/signup
