Google "Carbon" just won everything...

Summary

The video reviews Google’s announcement of Gemini 4 Carbon, which Google describes as its strongest model yet, with advanced cybersecurity and long-horizon capabilities.

The presenter says Carbon has not yet been released publicly. Google plans to provide it first to cybersecurity partners through its Fairwind program, continue testing and adjusting safeguards, and then gradually expand access to developers, businesses, and consumers.

Benchmarks and Capabilities

The presenter examines results from Google and Artificial Analysis’ Intelligence Index V5, where Carbon is reported to score 59, ahead of the models compared in the video. Other results highlighted include:

• Strong performance on coding and agentic tasks, including Terminal Bench 4.0.
• First-place results on several science and math benchmarks.
• A strong result on Vending-Bench, which tests a model’s ability to run a vending machine and maximize revenue.
• High cybersecurity scores, including on vulnerability-finding and exploit-related benchmarks. The presenter cautions that competitors’ public models may have some cybersecurity capabilities restricted.

The video also notes that Carbon performs well on computer-control testing but does not lead the cited OSWorld result: Claude Opus 5.5 ranks ahead of it.

Efficiency, Cost, and Alignment

Google lists introductory API prices of $2 per million input tokens and $10 per million output tokens. After an unspecified introductory period, those prices will rise to $4 and $20, respectively. Cached input tokens are said to cost 95% less than the standard input price.

The presenter emphasizes that the cost of completing a task depends on both token prices and how many tokens a model uses. Artificial Analysis figures cited in the video suggest that, on measured tasks, Carbon uses substantially fewer tokens than some competing models—about half as many as Sonnet 5.5 and Haiku 5.5, and roughly 20% fewer than Opus 5.5. However, the presenter says some OpenAI models remain more token-efficient and argues that OpenAI currently offers a strong range of quality-to-cost options.

Google also says Carbon reduces shortcut-seeking and cheating behavior, including attempts to exploit benchmarks, and makes fewer unsupported claims than its predecessor, Gemini 4 Argon. The presenter says these claims will need to be assessed once the model is available for testing.

Overall Assessment

The presenter is impressed by Carbon’s reported benchmark results and low pricing, and says Google’s progress makes him more optimistic about the company. However, because he has not yet used Carbon, his conclusions are based on published results rather than hands-on testing. Google’s staged rollout also means independent evaluation will have to wait.

Sponsor Feature Overview

The video includes a segment on Factory, a development platform whose AI agents are called “droids.” It can be used via CLI, desktop app, or Slack.

Factory’s “missions” feature divides larger projects among agents. A validator agent checks the resulting application and sends work back for iteration. Factory also has a model router; the sponsor claims it reduced costs by 63% in production compared with using frontier models directly.

Main Speakers and Sources

• Matthew Berman — presenter and commentator.
• Google’s Gemini announcement and blog — source of Carbon’s stated features, rollout plan, and pricing.
• Artificial Analysis — source of the benchmark index and token-usage comparisons discussed in the video.
• Factory — sponsor featured in the video.