Alphabet’s Gemini 4 clears pretraining, but the harder work is just beginning

Google announced the start of Gemini 4’s pretraining back on July 21, 2026, billing it as the largest pretraining effort in the company’s history. Now, as of early September, the model has shown strong preliminary results from that phase. The catch: posttraining hasn’t happened yet, and that’s where AI models go from impressive spreadsheet entries to products people can actually use. What Gemini 4 is actually being built to do CEO Sundar Pichai spelled out the ambition during Alphabet’s Q2 earnings call on July 23, framing Gemini 4 around two specific priorities: coding and autonomous agents. Pichai also noted that Gemini 4 would rely on what he called “much larger base models” to meet the demands ahead. The price tag reflects that. Alphabet has set its 2026 capital expenditure guidance as high as $205 billion, a figure that has pushed the company into negative free cash flow territory for the first time. Where the competitive pressure is coming from Gemini 3.5 Pro, the model that preceded this effort, reportedly hit delays tied to coding issues. Gemini 4 is, in part, a response to that. A model explicitly built around coding strength addresses one of the specific areas where compe...