Google Delays Gemini 3.5 Pro Again, Scraps Current Architecture

Google DeepMind has once again pushed back the launch of Gemini 3.5 Pro, this time to July 17. The company has abandoned its previous approach of building on the Gemini 2.5 Pro base and started a new pretraining cycle from scratch, which internal sources say is meant to close quality gaps exposed by testing with early users.

What's changing under the hood

Rather than stacking new layers on the proven Gemini 2.5 Pro base, the Google DeepMind team opted for a full rebuild of the model from the ground up. That kind of decision usually raises scheduling risk, since a new model has to be trained, tuned and tested almost from scratch, but it also opens the door to fixing structural limitations that patching the older architecture couldn't solve. Google wants to use the rebuild to improve the model's mathematical reasoning, SVG vector graphics generation, and overall image quality, areas where competitors have clearly been pulling ahead lately.

The model is also set to gain a reasoning layer called Deep Think, aimed at supporting more complex, multi-step tasks and autonomous workflows. Google is positioning Gemini 3.5 Pro for engineering and design use cases, where precision matters for generating code, interface mockups and 3D models.

Competitive pressure and researcher departures

The delay coincides with the launch of OpenAI's GPT-5.6 and Anthropic's Fable 5, adding further pressure on Google to make sure the new Gemini actually closes the gap with rivals rather than just playing catch-up. In the background, DeepMind has also been losing top researchers: in recent weeks the company saw departures including Noam Shazeer, co-author of the breakthrough transformer architecture, and John Jumper, a Nobel laureate for his work on AlphaFold, who moved to OpenAI and Anthropic respectively.

Markets reacted nervously to the news. On June 22, Alphabet shares dropped about 5 percent, equivalent to a roughly $225 billion drop in market capitalization, as investors weighed another delay to the flagship model against the loss of key researchers to rivals gearing up for IPOs valued above $100 billion.

What's next for Google's lineup

Alongside Gemini 3.5 Pro, Google's roadmap still includes a faster Gemini 4 Flash variant for simpler tasks and Nano Banana Pro, meant to compete with GPT-Image 2 in image generation. Another slip in the flagship model's launch date, though, is reason to treat these announcements with caution until the company backs them with firm dates.

For companies and developers using Gemini in Poland, this means, in practice, several more weeks of working on the 2.5 Pro architecture before the new model reaches even limited access on Vertex AI. Google is also positioning Gemini 3.5 Pro as a cheaper option for enterprise customers, so any further delay would hit hardest the companies that had planned a cost-driven migration this quarter.

Sources: BigGo Finance (finance.biggo.com), Startup Fortune (startupfortune.com), Investing.com Poland (pl.investing.com)