Closed Frontier Labs Are Pulling Ahead of Open AI | Melvin Johnson artwork

Closed Frontier Labs Are Pulling Ahead of Open AI | Melvin Johnson

MTS

November 13, 2026

Melvin Johnson, VP of research leading Gemini post-training at Google DeepMind, joins MTS to unpack the rapid rollout of the Gemini 4 model family—Argon, Carbon, and today’s workhorse release, Hydrogen.

0:00

Melvin Johnson

There's always going to be a lead with the frontier labs. Like, we're gonna not stop with hydrogen, we'll have the next element out, and we'll have the next frontier model out quickly. So it's always going to be a catch up game with open weights models. And if you look at the gap between closed and open weights models, it's the largest right now than it's ever been throughout this year. There were points in this year where the gap was much smaller, but right now the gap is quite large.

0:29

SPEAKER_2

And we are back. We are live with Melvin Johnson, who is Vice President of Research at Google DeepMind, where he leads Gemini post-training, which is pretty topical given all of the stuff that Gemini has been putting out over the last week. Melvin, welcome to MTS.

0:44

Melvin Johnson

Glad to be here. Thank you.

0:46

SPEAKER_2

So tell us about the whole Gemini 4 model family and how that came to be. You announced Argon, Carbon, and now today, Hydrogen.

0:55

Melvin Johnson

Yeah, I mean, it's been an exciting ride. You know, we've had all of these launches in the past month or so. So we started off with Argon, our first Gemini 4 frontier model. We quickly followed up with Carbon. I think, in fact, within the same month, we announced Carbon. And earlier this week, on Tuesday, we made Carbon generally available for everybody.

Between Argon and Carbon, there was a massive jump in both capability and better behavior. And today, we're launching Hydrogen, our next model in the Gemini 4 series.

Hydrogen is really going to be the daily driver and the workhorse model for everybody. And it's much faster, but delivering cutting edge capability. So we're really excited to roll it out to as many users and surfaces as possible.

And there's more to come in the Gemini 4 era, and we'll keep pushing models out.

1:58

SPEAKER_2

Awesome. So tell us more about Hydrogen specifically. I think I'm most curious about how it beats both Sonnet and Terra at Vending Bench.

2:11

Melvin Johnson

Yeah, I know. I mean, we're also very happy about that, but it's not surprising. If you look at the AI index score for Hydrogen, it actually beats Argon. Argon has an index score of 50, and Hydrogen at the high setting, the default setting is at 51, and at the extra high setting is at 54 So it's actually better than the frontier model that we released a month and a half ago. So it's not surprising that it's beating Sonnet and Terra at Vending Bench.

So it's a really good model.

2:52

SPEAKER_2

Totally.

What about Google's broader strategy around these models? Anthropic and Opening AI have been in a pretty clear lead for quite some time. It seems like Google is finally catching up with Carbon. What do you think it will take to durably keep Google in the lead?

3:14

Melvin Johnson

I think we have lots of friends that are unique to Google. We're vertically integrated with the chips, with DPU.

We really plan for training and inference on the DPUs, and we have a fast cycle with the chips that we have. We have 100,000 plus engineers using the models every day, giving us feedback. That's a lot of useful feedback. You really cut down the back and forth of having to test it with customers and users, and trying to translate that feedback back. You have people sitting in the building giving you feedback. That's a big advantage that we have, and we have massive distribution in all of our products. Gemini app is a billion plus product. We have AI mode and AI overviews that serves all of search users, and we have anti-gravity that is growing rapidly as the coding surface.

4:14

SPEAKER_2

Yeah, seems like AI overviews is seemingly the most used AI product in the world.

It is. It is.

4:23

Melvin Johnson

It is. And, you know, funny enough, I tell my researchers that work on AI overviews is like, hey, you're actually, you don't realize, you're building the most used AI model in the world by a big margin.

And, you know, we talk a lot about the frontier and, but only a small sliver of people have access to the frontier. But if you're distributing it to 7 billion plus people, that's a massive, that's a massive deal.

4:55

SPEAKER_2

How do you see hydrogen competing with open weight models? You know, there's always this narrative that open weight models will be able to erode the margins of the frontier labs because they will produce capabilities that are almost as good for substantially cheaper.

5:10

Melvin Johnson

Read the other 6 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000801711976