Google Delays Gemini 3.5 Pro Launch to July 17, Rebuilds Model From Scratch

Google DeepMind has pushed the launch of Gemini 3.5 Pro from June to July 17, 2026. The company scrapped the existing Gemini 2.5 Pro base architecture and started an entirely new training cycle for the successor model instead of simply fine-tuning the current one. The decision came just ahead of the launch of rival GPT-5.6 from OpenAI and coincided with the departure of two of DeepMind's top scientists to competitors.

Gemini 3.5 Pro was originally set to reach users in June, having been announced at the Google I/O conference in May. Instead, the DeepMind team decided to abandon the Gemini 2.5 Pro base and run a new pre-training cycle from zero, effectively building the model on a new foundation rather than iterating on the current version.

Architecture changes

According to industry reports, the delay stems from specific performance gaps Google wanted to close before launch: mathematical reasoning, SVG scene generation, and overall image generation quality. Rather than patching these shortcomings in the existing model, engineers opted for a full rebuild.

The new version is expected to offer context windows of up to 2 million tokens, a Deep Think reasoning layer built for multi-step tasks, and autonomous workflow execution features. Testing leaks also suggest Gemini 3.5 Pro will be stronger at generating interfaces, front-end code, and vector graphics, making it a more competitive tool for developers.

Departures coinciding with the delay

The delay decision coincided with the departure of two key DeepMind figures. On June 18, Noam Shazeer, co-lead of the Gemini project and co-author of the landmark 2017 paper 'Attention Is All You Need,' left for OpenAI. The following day, June 19, Nobel laureate John Jumper, creator of the AlphaFold system, joined Anthropic.

The market reacted nervously. On June 22, Alphabet shares dropped about 5 percent in a single session, equivalent to a loss of roughly $225 billion in market capitalization. Investors read the departures as a sign that competitive pressure from OpenAI and Anthropic was starting to cost Google its most valuable asset: its people.

Give us until next month - Sundar Pichai, CEO of Google and Alphabet, at the Google I/O conference on May 19

Racing GPT-5.6, Fable 5 and DeepSeek V4

The new launch date of July 17 puts Gemini 3.5 Pro up against China's DeepSeek V4, expected around the same time, as well as the newly released GPT-5.6 from OpenAI, whose Sol, Terra, and Luna models became widely available on July 9 following a safety review by the Trump administration. Google also wants its model to compete with Anthropic's Fable 5.

Industry analysts note that Google's strategy doesn't need to rely on outdoing rivals on parameter count, but rather on leveraging its advantage in breadth of world knowledge and integration with existing products such as Search, Workspace, and Android. Positioning Gemini 3.5 Pro as a cheaper but capable alternative is meant to attract companies sensitive to deployment costs.

What it means for businesses and developers

For tech teams using Google's models, the delay means the coming weeks are worth spending on testing the current 2.5 Pro version and preparing for migration rather than waiting idly for the successor. The promised improvements in front-end code and vector graphics generation could carry practical weight for companies building tools on top of the Gemini API.

The departure of top scientists to competitors is also a reminder that the fight for talent in the AI industry directly affects the pace of development of products used by companies worldwide, including in Poland. The July 17 launch will be a test of whether Google has genuinely made up for lost time, or whether the delay will show in the quality of the model's first public release.

Sources: TechTimes (techtimes.com), HackerNoon (hackernoon.com), Geeky Gadgets (geeky-gadgets.com), AiBase (news.aibase.com).