Short answer: Jordan Metzner says Gemini 3 impressed him for coding and design taste despite throttling in Antigravity. Carina says it is cheaper to run than GPT, scored 100% on AIME 2025, and Axiom's model beat it on one benchmark.
Watch this answer from Episode 22 (starts at 18:01) · Episode 22 notes and transcript
From Episode 22: Let's do the Math!, starting at 18:01.
Jordan Metzner [18:01]: Thanks for joining us. Let's talk a little bit about the news for this week as it's been another crazy week. Gemini three just dropped. I don't know if you've tried it yet. I downloaded Antigravity, the new Google code editor. Lots to tell you guys, but I'm being throttled everywhere. So Cursor, I can't get Gemini three to work. I can't get Gemini three to work in Antigravity. I burned all my tokens. So, I'm I'm munching tokens all night long pushing Gemini three, but I've been incredibly impressed with its performance for coding and design taste, also much nicer.
Carina [18:33]: We like Gemini three. It's also, like, so much cheaper to run compared to, like, GPT. One thing that's very interesting is we you know, on a certain benchmark, I can tell, which one, and our model actually beats Gemini three
Jordan Metzner [18:47]: Oh, wow.
Carina [18:48]: In math improves yesterday.
Sam Nadler [18:50]: I know Gemini three has been outperforming almost every LLM on the market in terms of test scores. I tried to use it quickly this morning to build a personal meditation app, and I bumped into some hiccups, but maybe it was user error and not the model.
Carina [19:08]: It got 100% on AIME 2025, which is the high school level. So it's three levels before the IMO. So if you win AIME, you will go to US AMO, and then you'll go to MOP, and then you will eventually go to IMO. But it got 100% on Yeah. Yeah. That's fascinating. But 100% does code execution shows how important the
Jordan Metzner [19:29]: Code is needed. Yeah.
Carina [19:30]: Yeah. That's the tool is important.
Jordan Metzner [19:32]: Well, It says without tools, it got 95%. So, obviously, that's a significant improvement.
Carina [19:38]: Odd because it has what? The it has 20 well, guess that's odd because it has, like the number of AIME problems is 25 problems?
Jordan Metzner [19:49]: Well, if you compare it to Gemini, looks like to
Carina [19:51]: Two Fifteen. Point 15 problems. So 15 problems on the AIME exam, so 95% is, like you know, got one wrong.
Jordan Metzner [20:00]: Yeah. So they got a 88 last time, so they probably got two wrong without the tool. You know? Yeah. Mhmm. Pretty cool. But maybe it's calculated differently because Sonic got a 80 If it got a 88.
Carina [20:15]: I guess it's like, past Yeah. Probably. Yeah. Past k.
Jordan Metzner [20:20]: Okay. But anyway, yeah, Gemini. Super cool. So, I just think this is interesting. The Gemini model for a lot of reasons. I mean, one, this is the first, real big model coming out of, one of the type three hyperscalers, know, the Amazon, Google, Microsoft. So far, all those models have been behind, pretty much Claude and Anthropic and OpenAI historically, and this is a big leap. Google has a lot of money and a lot of servers and a lot of warehouses and a lot of products to inseminate, Gemini three Pro into their products from, Google Suite to search to YouTube to ads. I mean
Carina [20:54]: Yeah.
Jordan Metzner [20:55]: I don't know. Big opportunity.
Sam Nadler [20:57]: Yeah. 650,000,000 monthly active users across search, the Gemini app. I mean, that's an insane amount.
Jordan Metzner [21:05]: Yeah. Insane amount of users. Obviously, we saw this week Warren Buffett made a massive investment in Google. And I think if you saw the videos of I don't know if you guys saw the videos from Antigravity, the launch of the code editor, Sergey was front and center, you know. So yeah.
