Google Gemini 3.5 Pro Flagship Model Delayed by Several Months
Google's Gemini 3.5 Pro, originally scheduled for release in June, has been delayed due to its coding capabilities not meeting internal targets. Even after updating training data in late June, the results remain unsatisfactory.
Insiders revealed that the latest evaluations may even fall short of Meta's Muse Spark, and some employees in related business areas have had their testing permissions revoked, with only the model team allowed to continue using and viewing results.
The Gemini team is demoralized, with some employees concerned about integration and layoff risks. Notably, Noam Shazeer has moved to OpenAI, and John Jumper has joined Anthropic. Google responded that the model is still being tested with partners and that the Flash version is being upgraded.
Source: Public Information
ABAB AI Insight
Since Gemini 1.0, Google has iterated its flagship model multiple times. The delay of 3.5 Pro continues its internal pursuit of high standards for coding capabilities, similar to the adjustments made during the transition from Bard to Gemini.
Google's internal revocation of testing permissions and the risk of talent loss aim to focus on core team challenges, reflecting the pressure in the large model competition for resources to concentrate on rapidly iterating laboratories.
With OpenAI and Anthropic attracting key talent from Google to accelerate their own progress, and the impact of competitive models like Meta's Muse Spark, Google is currently at a critical stage of parallel testing of multiple models and talent retention.
This essentially involves talent and technology substitution: top researchers are moving to high-execution laboratories, and Google is responding to capability gaps through delays and internal controls, while industry pricing power is concentrating on models that can quickly validate hard metrics like coding, accelerating resource reorganization among AI laboratories.
ABAB News · Cognitive Laws
Delays in flagship models often expose the gap between internal goals and reality, leading to talent movement.
In the large model competition, coding capability is the moat; falling behind in evaluations puts both morale and resources under pressure.
Top talent chases optimal execution environments, and the success of laboratories is determined by validation speed rather than historical brand.