Google is reported to have delayed the release of Gemini 3.5 Pro as it continues to improve model programming capabilities
The release of Gemini 3.5 Pro, the most powerful flagship AI model, has been delayed several months from the original plan as the company invested more time in improving model capabilities, especially programming capabilities. According to 10 current and former employees, the delayed release of Gemini 3.5 Pro has caused dissatisfaction among Google engineers, AI researchers, and management. They are worried that as Anthropic and OpenAI successively launch models that surpass Gemini in performance, Google may lose its market lead. The above-mentioned sources said that Google needs to go through multiple stakeholder approvals before the model is released, and it also needs to integrate AI capabilities into products such as search, maps, and YouTube. These factors will slow down the product release progress. Relevant people requested anonymity due to internal company circumstances. OpenAI and Meta Platforms Inc. Both have recently released new models that further lead Google in terms of programming capabilities. A person familiar with the matter said that Google updated the data used for Gemini training late last month in the hope of improving its programming capabilities, but the results were disappointing. Google shares fell 3.2% on Thursday. A Google spokesperson said in a statement: "We are accelerating the launch of multiple types of models while maintaining high cost-effectiveness." The spokesperson said that Google is in communication with the US government on Gemini capabilities and AI safety standards. "We are currently testing 3.5 Pro, upgraded Flash models, and other models with our partners, and are actively working with the U.S. government on model testing and broader regulatory framework discussions," the statement said. At the beginning of this year, Anthropic had to temporarily withdraw related products under pressure from the U.S. government after internal testing found that its latest model had dangerous cybersecurity capabilities. OpenAI has also voluntarily restricted and phased the launch of its latest AI models due to national security concerns and pressure from the Trump administration. Google's vast product ecosystem provides an entry point for ordinary users to access generative AI, and they can also use the data accumulated by these products to improve model performance. But one former employee said getting business units to move in the same direction was "like boiling the entire ocean." Current and former employees say maintaining a unified strategy becomes more difficult when a company's strategy is constantly changing or when multiple departments are doing the same work. At the same time, this also makes it more difficult for a single project to obtain sufficient resources and market attention. People familiar with the matter said that Google Cloud, Google DeepMind and Android teams are all developing AI programming tools for developers, and some consumer product teams are also involved. Former employees said that some engineers within Google held a more conservative view and believed that all important code should be written by humans to comply with Google standards. They also said that when the AI tool was first launched, employees were restricted from using Gemini to write or analyze software code because the company was worried that proprietary code would be leaked into model training data. Although these regulations have since been relaxed, they early limited opportunities for engineers to try out AI development. Google said that the company had disclosed at the recent Cloud Conference that currently about 75% of the code is generated by AI and put into the production environment after manual review, which meets Google standards. Today, Google engineers are encouraged to use AI to generate code. However, employees said that in actual use, they often encountered capacity constraints due to competition for computing resources within the company. AI researchers say that Gemini's biggest advantage is its ability to call Google search data, and Anthropic and OpenAI currently take the lead in building the strongest AI models. Former Google employees said that some researchers have jumped to leading AI labs such as Anthropic because they are disappointed with Google's situation in the AI race. Currently, only some teams within Google can use Anthropic’s Claude model. Google has restricted access to the tool to teams working on cutting-edge research and other high-priority projects.
While waiting for the release of Gemini 3.5 Pro, Google customers have received mixed reviews about Gemini 3.5 Flash. Rodrigo Davies, product manager of the design platform Figma, said that the company has recently connected 3.5 Flash to the newly launched "Figma Agent" AI assistant to help designers generate and iterate ideas. For Figma, this model achieves the best balance between speed and quality However, Freddy Vega, founder and CEO of Platzi, a Latin American education technology platform, said that 3.5 Flash is in an embarrassing situation. It is more expensive than the 3.1 Flash previously launched by Google, but it is slower and its capabilities are still significantly inferior to competitors' high-end models. He said that the team has now switched to Anthropic's mid-range model to balance reasoning capabilities and response speed.
While waiting for the release of Gemini 3.5 Pro, Google customers have received mixed reviews about Gemini 3.5 Flash. Rodrigo Davies, product manager of the design platform Figma, said that the company has recently connected 3.5 Flash to the newly launched "Figma Agent" AI assistant to help designers generate and iterate ideas. For Figma, this model achieves the best balance between speed and quality However, Freddy Vega, founder and CEO of Platzi, a Latin American education technology platform, said that 3.5 Flash is in an embarrassing situation. It is more expensive than the 3.1 Flash previously launched by Google, but it is slower and its capabilities are still significantly inferior to competitors' high-end models. He said that the team has now switched to Anthropic's mid-range model to balance reasoning capabilities and response speed.