# Google Delays Gemini 3.5 Pro as Coding Improvements Fall Short > Source: > Published: 2026-07-17 16:49:56+00:00 Google has postponed the release of its flagship Gemini 3.5 Pro AI model after internal efforts to improve its coding capabilities reportedly failed to meet internal performance goals, a move that puts Google at risk of losing ground to rivals like OpenAI and Anthropic. Gemini 3.5 Pro was expected to reach customers shortly after Google introduced the model during its I/O developer conference in May. Instead, the release has slipped by several months while engineers work to improve the model, particularly its ability to generate software code, according to multiple reports citing people familiar with the project. Google has maintained that development is progressing. A company spokesperson said it is testing Gemini 3.5 Pro, an upgraded Flash model and additional AI systems with partners while continuing discussions with the US government on model testing and broader AI safety frameworks. The company also said it remains focused on delivering models that balance performance with cost efficiency. The postponement demonstrates the highly demanding standards frontier AI developers face before releasing new foundation models. Beyond improving benchmark scores, companies must demonstrate reliability and address growing government scrutiny around security before debuting new versions of leading systems. “Google delaying Gemini 3.5 Pro after missing its own coding targets confirms coding has become the deciding test of frontier model credibility. Scrapping the base model for deeper retraining shows Google grading itself against shipping competitors, a harder bar than its own roadmap,” Mitch Ashley, VP, Software Lifecycle Engineering at the Futurum Group, told Techstrong.ai. “Enterprises building coding agents on frontier models should treat vendor roadmap dates and benchmark claims as provisional until verified in their own pipelines. Teams architecting around a single vendor’s coding trajectory carry a dependency they cannot control.” **The AI-Generated Code Market** Google’s challenge is particularly significant because AI-generated software development has become such a competitive sector. Meta recently introduced Muse Spark 1.1, which the company describes as its strongest model yet for coding and autonomous AI agents. OpenAI launched GPT-5.6 Sol, claiming the model delivers 54% greater token efficiency on agentic coding workloads. Internally, Google has been working to consolidate AI development efforts across several engineering organizations while expanding AI-assisted programming inside the company. Google says AI now generates roughly 75% of the code that ultimately passes review and reaches production. It has also begun unifying coding infrastructure under its Google Antigravity platform, and has created a dedicated AI coding team within DeepMind. Even with those investments, news reports indicate that competing priorities across Google Cloud, DeepMind, Android and consumer product teams have complicated development. Researchers have also cited shortages of computing resources and internal restrictions that previously limited experimentation with AI-generated code. Some researchers have left Google for competitors including Anthropic and OpenAI. Bottom line here, the postponement reveals the many challenges in the fiercely competitive market for AI models. Enterprises, now with more experience in AI development, are evaluating systems on coding performance, cost efficiency, reliability and their ability to support increasingly sophisticated AI agents, forcing model developers to a higher standard.