New Leading AI Models as Benchmarks Shift
Google's Gemini 4 Carbon took first place on Artificial Analysis's rebuilt Intelligence Index V5, which shows that benchmarks must keep changing to track real AI use, even though they stay limited.
Watch on YouTube ↗Charlie's summary
- ▶ 1:00
Google's quick Gemini releases come from iterating with reinforcement learning on an existing base model, which puts it back in the frontier race
Gemini 4 Carbon arrived just two weeks after Argon. A new base model lets Google iterate with reinforcement learning instead of waiting on a lengthy pre-training run. Carbon topped six of eight key benchmarks and scored 59 on Index V5. Sundar Pichai expects acceleration over the next 12 months. The host suspends judgment until the model is actually released, since access is limited to cyber defenders.
- ▶ 6:07
Google's cloud growth is strong, but AI Mode's lower monetization is eating into search growth
Google Cloud doubled revenue to $30.7 billion, nearly closing the gap with Microsoft Azure. Search growth slowed to 12%, from 17% and 19% in the prior two quarters. Pichai acknowledged that strong AI Mode query growth hurts because of its lower monetization. Some analysts see Google reaching the point where it must cannibalize its own search business. The stock fell as much as 10% after hours.
- ▶ 8:09
The CapEx slowdown looks like physical constraints, not weak demand, so it does not fit the bubble-popping theory
Google and Microsoft held CapEx forecasts steady, but actual spend came in under expectations. Google guides to $300 billion next year, yet is undershooting this year's numbers. Bears had predicted a CapEx slowdown would signal faltering demand. The host argues that hyperscalers would keep accelerating if they could, and are instead limited by components, power, labor and community support.
- ▶ 14:48
Benchmarks are inherently limited but important, and Artificial Analysis keeps retuning its index to match real use
The host says benchmarks are the only alternative to pure vibes, but no single score should be treated as gospel. Artificial Analysis has been in a cat-and-mouse game: V4.1 underweighted agentic work, so GPT-6 Astra scored poorly against people's lived experience. The index was adjusted to track how models are actually used. The host's advice is to look at individual benchmarks for the use cases you care about.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.