Summary
Developer and YouTuber Theo returns to Google’s Gemini 4 Carbon with a harsher verdict than in his earlier video, arguing that he underestimated how poor the model is for real-world engineering work.
After auditing his actual usage across models, Theo says Gemini 4 Carbon was consistently about five times slower than competing models on comparable tasks, not merely slower as he first reported. He also criticizes the cost of the subscription subsidy and argues that Google’s inference hosting does not match the standards of other labs. He says the only clear advantage he found is that it is slightly less likely to flag cybersecurity-related requests.
Theo compares the situation to an earlier episode in which Azure’s hosting of OpenAI models had occasional first-token delays of 100 to 200 seconds, which was eventually fixed after he raised it publicly. He says he hopes the video helps teams at Google DeepMind push for fixes, while acknowledging that many people there are responsive and open to feedback.
The video is a pointed look at the gap between benchmark claims and day-to-day developer experience with frontier models, including speed, reliability, and pricing.
📺 Source: Theo – t3․gg · Published November 13, 2026
🏷️ Format: Opinion Editorial
