Google DeepMind Delays Gemini 3.5 Pro Release to July 17 for Native Gemini 3 Retraining

Google DeepMind confirmed on July 6 that Gemini 3.5 Pro's public launch slips to July 17, 2026. The stated reason is a foundational rebuild: rather than extend the existing Gemini 2.5 Pro base layer, DeepMind is running an extended pre-training cycle on a native Gemini 3 foundation. The goal is to lift performance in the areas where internal evals reportedly show gaps — multi-step mathematical reasoning and SVG scene generation among them.

The delay is a tacit acknowledgment of how fast the frontier is moving. In the same week, Anthropic shipped Claude Sonnet 5 (approaching Opus 4.8) and OpenAI previewed GPT-5.6 Sol with strong coding and scientific reasoning. A 3.5 Pro built on last-generation foundations risked launching already behind, so DeepMind is trading a two-week slip for a stronger base.

Crucially, Google is not going dark: Gemini 3.5 Flash — which shipped July 2 with computer-use capabilities — remains available, as do Nano Banana 2 Lite and Gemini Omni Flash, and a Gemini 2.5 Pro API update added structured JSON output and improved long-context handling. So the delay is specific to the flagship Pro tier, not the whole line.

What to watch: whether July 17 holds, and whether the native-Gemini-3 base actually delivers the reasoning wins DeepMind is promising. Delays framed as quality moves can also read as competitive pressure — investors and developers will judge 3.5 Pro against Sol and Fable 5 on the multi-step reasoning benchmarks Google itself flagged as the reason for the wait.