Google’s Gemini Pro slips as coding goals miss target

Google is months behind schedule on the next version of Gemini Pro, with the delay tied to coding performance that has failed to meet internal goals, according to a Bloomberg report citing 10 current and former employees. The upgrade had been widely expected at Google’s May developer conference, but the company has not closed the gap with Anthropic and OpenAI, whose latest models are said to outperform Google’s current systems at writing code.

The report rattled investors: Alphabet shares fell more than 3% on the news. Bloomberg says Google updated Gemini’s training data late last month to improve coding, but one source described the results as disappointing. Pressure has only increased after OpenAI and Meta released newer models that reportedly extend their lead in code generation.

A Google spokesperson told Bloomberg that the company is “shipping quickly across a wide range of models” and is testing the upgraded Pro, a new Flash model, and other models with partners.

Google’s internal AI coding push

Part of the slowdown appears to be organizational. Google Cloud, DeepMind, and the Android team are all building AI coding tools, with consumer product teams involved as well, creating internal competition that has slowed progress.

According to former employees cited by Bloomberg, Sergey Brin has pushed the company to move faster on AI coding, but those efforts have run into competing internal factions and resistance from engineers who believe important code should still be written by humans to meet Google’s standards.

Recommended reading

Hassabis says STEM makes you 10x better at AI

Google has started to consolidate those efforts. Chief AI Architect Koray Kavukcuoglu is working to unify internal coding tools, and a new DeepMind team led by research engineer Sebastian Borgeaud has been set up specifically for the problem. At its most recent Cloud conference, Google said 75% of code at the company is now AI-generated and that it has consolidated most developer tooling under Antigravity, its internal platform for data, memory, and safety protocols.

The delays have also fed a wave of senior departures to Anthropic and other labs, with former employees saying frustration over Google’s competitive position is one reason. Engineers inside Google reportedly also face capacity limits when trying to use AI tools, due to internal competition for compute resources — an issue that also affects external customers.

Only some Google teams are allowed to use Anthropic’s Claude, with access limited to groups doing cutting-edge research. Among customers, reaction to the current Flash model has been mixed. Rodrigo Davies, a product manager at Figma, said it hit “a sweet spot of speed and quality” for the company’s AI assistant. But Freddy Vega, CEO of Latin American education platform Platzi, said Flash is more expensive and slower than its predecessor while still being far less capable than competitors, and said his team has moved to Anthropic instead.