Media: Google is unhappy with Gemini 4's capabilities ahead of its launch
Inside Google, some employees are skeptical about the capabilities of the new flagship artificial intelligence model Gemini 4, despite high results in industry tests. In practice, the model, in particular, faces difficulties when performing some programming tasks.
This was reported by Bloomberg, citing employees with direct access to the development.
According to the agency, Gemini 4 performs well in standard benchmarks used by technology companies to measure the capabilities of new AI models. However, some Google employees believe that results transfer worse to real-world work tasks, especially those related to writing and analyzing software code.
Google itself disagreed with this characterization. Company representatives told Bloomberg that it would be wrong to talk about weak results from Gemini 4 in programming.
One employee familiar with the development also told the agency that there is a "broad consensus" within the company that Gemini 4 is at the cutting edge among modern AI models. At the same time, some specialists believe that competing developments from OpenAI and Anthropic have been improving faster recently.
Doubts around Gemini 4 are especially sensitive for Google after problems with the previous flagship model. The company planned to release Gemini 3.5 Pro back in the summer, but ultimately canceled the launch. It was previously reported that the model failed to meet Google's internal goals, including in programming.
At Google I/O in May, the company said Gemini 3.5 Pro was already being tested inside Google and was supposed to be released the following month. Instead, the developer focused on faster and cheaper models in the Flash family.
In July, Alphabet CEO Sundar Pichai confirmed that Google had begun training Gemini 4, calling it the largest pretraining cycle for a model in the company's history.
Last week, new Google DeepMind head Koray Kavukcuoglu said Gemini 4 has already moved to the post-training stage — the final tuning after the main training process. According to him, the development results are encouraging for the team and Google expects to bring the model to market significantly before the end of the year.
Gemini is particularly important for Google because models in this family are being gradually integrated into almost all of the company's biggest products — from search and Gmail to Chrome, Android, and cloud services. Therefore, Google's position in competition with OpenAI and Anthropic depends on the capabilities of the new generation.
