I was wrong...

Summary

The creator revisits his earlier review of Gemini 4 Carbon, saying he was wrong to understate how slow it was. After auditing comparable real-world coding-agent tasks, he reports that it was consistently about five times slower than the other models he tested. He says his estimate may even understate the gap because he had instructed his agents to stop waiting for Gemini when they had other work to do.

His central technical criticism is that the delay is not just a matter of token-generation speed. He says Gemini often takes more than 11 seconds to begin responding after a request reaches Google’s server. In agent workflows, where a task involves multiple requests and tool calls, that startup delay accumulates. He also cites inefficient tool use, repeated file reads, and unreliable prompt caching. In his tests, cache reads sometimes failed to use the cache that had just been written or used an older cache instead.

The creator argues that standard benchmark comparisons can miss these problems: many benchmark tasks are single-turn, while coding agents make repeated, multi-step requests. He also criticizes the hosting and API experience, including what he describes as confusing model-name conventions and differences between available endpoints. He says Google’s explanation that latency involves a trade-off with cost does not address his concern, particularly since he considers the model both expensive and slow. He also says its price was due to double on December 1.

Although he acknowledges that Gemini may be less likely to flag some cybersecurity-related prompts and has a few other behavioral differences, he concludes that it is not worth using for routine engineering work in its current state. He recommends that developers use alternatives instead and says he hopes Google will improve the model and its supporting infrastructure. He notes that people at Google DeepMind had been responsive to feedback, while criticizing one leader’s public response to the performance complaints.

Sponsor Segment

The video is sponsored by Infinite Red, a React Native consultancy. The creator describes its help with mobile-app architecture and React Native, as well as native technologies such as Swift, Kotlin, Objective-C, and Java. He argues that well-structured, “agent-friendly” codebases can help both developers and AI coding agents avoid common mobile-development pitfalls.

Main Speakers and Sources

• Theo, the video’s host and reviewer, provides the performance findings, comparisons, and opinions.
• Google DeepMind staff and a leader identified in the subtitles as “Verun/Arun” are referenced through reported interactions and public comments.
• Infinite Red is featured in the sponsored segment.