Google Employees Question Gemini 4’s Real-World Performance
Some Google employees have reportedly questioned whether Gemini 4’s strong benchmark results accurately reflect its performance on practical tasks, particularly software development.
Google is facing internal debate over the performance of its upcoming Gemini 4 AI model, with some employees reportedly questioning whether the system performs as strongly in real-world applications as it does on widely used industry benchmarks. According to reporting based on people familiar with the project, Gemini 4 has produced strong results in formal evaluations, but some Google employees have encountered inconsistencies when using it for practical coding and other complex tasks.
One area receiving particular attention is software development. Some employees reportedly found Gemini 4 less consistent when handling certain coding tasks, including front-end development. The concerns illustrate the difference that can exist between benchmark performance and everyday use, where AI systems must deal with ambiguous instructions, changing requirements and longer sequences of tasks rather than isolated evaluation problems.
The views inside Google are not uniform. Some employees reportedly believe rival models from companies such as OpenAI and Anthropic are improving at a faster pace in certain areas, while others believe Gemini 4 has reached the frontier of AI capabilities. A Google employee familiar with the model’s development reportedly disputed the criticism and said there was broad internal agreement that Gemini 4 was performing at the frontier, adding that Google had conducted extensive testing of the system.
Google has also pushed back against the characterization that Gemini 4 is underperforming at coding. The company has pointed to positive internal evaluations and comments from Google DeepMind leadership about the model’s development. Gemini technology is particularly important to Google because its models are integrated into major products and services, including Search, Maps, Gmail and Chrome.
The debate comes as Google intensifies its competition with other AI developers. Gemini 4 Argon has been positioned for complex tasks including software engineering, enterprise workloads and cybersecurity, while Google is also expanding its use of AI agents in model development and security testing. The company recently confirmed that experimental Gemini systems accidentally accessed three real companies during a cybersecurity test after an evaluation environment was misconfigured, highlighting the increasingly complex challenges involved in testing powerful AI systems.
