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

Google has pushed the launch of Gemini 3.5 Pro back to July 17 and pulled the already-finished model architecture from the production line at the last minute. The company decided to abandon the Gemini 2.5 Pro base and rerun the entire pretraining cycle from scratch, meaning weeks of additional work just before the planned launch.

Sundar Pichai had already teased Gemini 3.5 Pro as the next step in the Gemini model line at June's Google I/O conference, suggesting a launch within the month. DeepMind engineers reportedly pulled the architecture built on Gemini 2.5 Pro from production at the last moment, judging it inadequate given the pace at which rival models are advancing.

Why the Rebuild Happened

According to industry reports, the finished base was scrapped over concerns about quality regressions in key tasks, particularly mathematical reasoning, SVG vector graphics generation, and image quality. Rather than patch the existing architecture, the team opted for a costly but more thorough step: a new pretraining cycle from zero.

It's an unusual decision at such a late stage of model development, where fixes are typically cosmetic. It suggests that internal tests of Gemini 3.5 Pro were underperforming expectations, especially against the latest OpenAI and Anthropic models, which have repeatedly raised the bar on reasoning and coding benchmarks in recent weeks.

New Capabilities

The rebuilt Gemini 3.5 Pro is expected to offer a context window expanded to 2 million tokens along with a new reasoning layer called Deep Think, designed for solving complex, multi-step problems. The model is also set to gain expanded autonomous task execution capabilities, including working with external tools and code without constant user supervision.

Google is positioning the model differently than its rivals: rather than competing purely on raw performance in the premium segment, Gemini 3.5 Pro is expected to be pitched as a cheaper-to-run alternative for companies deploying AI at scale. Alongside it, Google is also expected to launch Gemini 4 Flash, a variant built for response speed, and Nano Banana Pro, an image-generation model competing with OpenAI's GPT-Image 2.

Three Labs Racing

Google's delay coincides with the launch of GPT-5.6, which OpenAI rolled out to the wider public on July 9, literally one day before news of the Gemini rebuild broke. The coincidence shows just how tight the industry's launch calendar has become: any delay from one player is immediately read as weakness against rivals who happen to be announcing their own news at the same time.

The stakes are high for Google. The company has already been grappling in recent weeks with a wave of departures of top DeepMind researchers to Anthropic and OpenAI, which some commentators read as a sign of a weakening position in the race for the best model on the market. The launch delay, even if justified by quality concerns, feeds into that narrative.

What It Means for Companies in Poland

For companies and developers using Gemini in Poland, this means a few extra weeks working with the older version of the model before new features become available, including the much larger context window useful for analyzing long documents or large codebases. Decisions to migrate Gemini-based integrations are best held off until the July 17 launch, to avoid duplicating work when adapting to the API.

Sources: Google Delays Gemini 3.5 Pro Launch to July 17 for Full Architectural Rebuild (finance.biggo.com), Gemini 3.5 Pro Targets July 17 as DeepSeek's July 24 Deadline Hits Developers Now (techtimes.com)