[Google launches Gemini 4 to catch up with AI rivals, with mixed internal evaluations of its programming performance] On October 1, Google has begun to launch the much-anticipated flagship artificial intelligence model Gemini 4 Argon, but there are questions within the company about the actual performance of the model in key areas such as programming. On Wednesday, Google opened the model to a small group of trusted cybersecurity partners and said it would expand its use after more testing, starting with paying users. Gemini 4 achieved leading results in multiple benchmarks, including beating OpenAI's Astra model in a test that measured security capabilities, the company said. However, some insiders say these metrics don't tell the full story. While Gemini 4 performed well in benchmark tests that measure model performance, it didn't perform as well when employees actually used it, especially for certain programming tasks, according to people with direct knowledge of the work.