Google helped kickstart the modern AI race, but staying ahead has turned out to be far more difficult than joining it. According to a new Bloomberg report, the company has fallen months behind its internal schedule for launching Gemini 3.5 Pro, its next flagship AI model, as engineers continue working to improve one of its biggest weaknesses: coding.
The delay isn’t simply about polishing another chatbot. It highlights a broader problem facing Google, where massive engineering teams, multiple product divisions and increasingly strict AI safety requirements are slowing the company’s ability to respond to rivals that seem happy to move much faster.
While OpenAI, Anthropic and Meta continue releasing increasingly capable models, Google appears to be caught in the difficult balancing act of building better AI without breaking the trust it has established across products used by billions of people.
Coding remains Gemini’s biggest challenge
Bloomberg, citing multiple current and former Google employees, reports that Gemini 3.5 Pro has been delayed because the company hasn’t achieved the improvements it expected in coding performance. The report says Google even refreshed the model’s training data late last month to boost coding capabilities, but the results reportedly failed to meet internal expectations.
That is becoming an increasingly important battleground. Writing code has emerged as one of the clearest benchmarks separating today’s leading AI models. OpenAI, Anthropic and, more recently, Meta have all invested heavily in developer-focused AI systems that can write, debug and reason through complex software projects. According to the report, both OpenAI and Meta currently outperform Google’s available models in this area.
Google, however, insists development is progressing. In a statement cited by Bloomberg, the company said it is testing Gemini 3.5 Pro, an upgraded Flash model, and other AI systems with partners while continuing discussions with the US government around testing standards and AI safety.
The delay is also notable because many observers expected Gemini 3.5 Pro to debut during Google I/O earlier this year. Instead, the company focused on incremental Gemini improvements while competitors continued shipping new frontier models.
Google’s biggest strength may also be slowing it down
Unlike most AI startups, Google isn’t building models in isolation. Every major Gemini release eventually needs to work across Search, YouTube, Maps, Android, Workspace, Cloud and dozens of other products. That scale gives Google enormous advantages, including access to unmatched amounts of real-world data, but it also introduces layers of internal coordination that can slow decision-making.
According to Bloomberg’s report, current and former employees describe competing priorities across DeepMind, Google Cloud, Android and other teams, with overlapping AI coding efforts making it harder to maintain a unified strategy. Former employees also said internal disagreements over AI-generated code and earlier restrictions on using Gemini for software development limited experimentation during the technology’s early rollout.
Google says those policies have evolved. The company claims roughly 75 percent of its production code is now generated using AI, while internal coding tools are being consolidated under a common platform called Google Antigravity. The report also notes that engineers are now expected to use AI for coding, although some continue to face computing capacity constraints due to intense internal demand for GPU resources.
The report also points to growing frustration within parts of Google’s AI organisation, with some researchers reportedly leaving for competitors like Anthropic. Meanwhile, customers appear divided on Gemini 3.5 Flash. While companies like Figma have praised its balance of speed and quality, others, including education platform Platzi, reportedly believe it sits in an awkward middle ground, offering higher costs than previous Flash models without matching the reasoning capabilities of premium rivals.
The bigger picture is that Google’s AI challenge is no longer about proving it can build frontier models. Few doubt that it can. The real question is whether a company of Google’s size can ship those models quickly enough in an industry where competitors now measure progress in weeks instead of months.