Google Reportedly Delays Gemini 3.5 Pro to July 17, Scrapping Its Base Model

According to reports, DeepMind threw out the Gemini 2.5 Pro-based foundation days before launch and started a fresh pre-training run — leaving Google the only major lab without a public flagship as GPT-5.6 and Grok 4.5 ship.

Jul 9, 2026Source: Geeky Gadgets

Google Reportedly Delays Gemini 3.5 Pro to July 17, Scrapping Its Base Model

Key Points

1.Reports say Google DeepMind pushed Gemini 3.5 Pro from a June target to roughly July 17 and scrapped the base model built on the Gemini 2.5 Pro architecture in favor of a ground-up rebuild; Google has not officially confirmed a date and a spokesperson declined to comment.
2.Leaked internal evaluations cited by the reports flag weaknesses in multi-step mathematical reasoning, recursive tool-calling and structural consistency in complex layouts and SVG generation; the rebuild reportedly targets a 2-million-token context window and a 'Deep Think' reasoning layer.
3.The delay lands during a talent exodus — Noam Shazeer left for OpenAI and AlphaFold's John Jumper plus two other researchers left for Anthropic in June — and leaves Gemini 3.5 Pro in limited Vertex AI preview while every rival ships publicly.

Google's next flagship is slipping again, and this time the reason reportedly runs deeper than polish. According to multiple reports, Google DeepMind has delayed Gemini 3.5 Pro from a June target to around July 17 and thrown out the model's base — the foundation built on the Gemini 2.5 Pro architecture — to start a fresh pre-training run. Google has not officially announced a date, and a spokesperson declined to comment on the timeline.

What 'scrapped base model' means

Rather than incrementally refine the 2.5 Pro foundation, DeepMind reportedly abandoned it outright — pulling the architecture from production pipelines days before a targeted deployment and opting for a ground-up redesign. The trade is a near-term delay in exchange for what the reports describe as a "deeply upgraded foundation."

The reported reasons

Leaked internal evaluations, as characterized by the reports, showed the scrapped model struggling with recursive tool-calling, faltering on multi-step mathematical reasoning, and failing to hold structural consistency when generating complex, multi-layered layouts and SVG scenes — areas where rivals have grown more stable. The rebuild reportedly aims at a 2-million-token context window, a "Deep Think" reasoning layer for multi-step problems, and stronger image and front-end generation. Every one of these specifics remains unconfirmed by Google; treat them as leaks, not launch notes.

A brutal month for the bench

The delay compounds a rough June for DeepMind's roster. Noam Shazeer, a Gemini co-lead and co-author of the 2017 "Attention Is All You Need" paper, left for OpenAI; a day later, Nobel laureate and AlphaFold creator John Jumper departed for Anthropic, followed by researchers Jonas Adler and Alexander Pritzel. Demis Hassabis has pushed back on the narrative, telling Semafor DeepMind still has "by far the biggest and broadest research bench of any of the labs out there."

Alone in preview

The timing is unforgiving. With OpenAI's GPT-5.6 rolling out publicly on July 9 and SpaceXAI's Grok 4.5 already live, Gemini 3.5 Pro remains confined to a limited Vertex AI enterprise preview — making Google the only major frontier lab without a publicly available new flagship. Each day in preview after the GPT-5.6 launch compounds a positioning problem Google can't benchmark its way out of until the model actually ships.

Why It Matters

Ripping up a base model days from launch is an expensive admission that catch-up patches weren't closing the gap on reasoning and tool use — the exact skills that decide agentic workloads. If the rebuild lands on July 17 and delivers, the delay reads as discipline; if it slips again amid the talent drain, it hardens a narrative that DeepMind is losing the frontier race it once defined.

Filed underGoogle AI AI Models #Demis Hassabis #Gemini 3.5 Pro #Google #Google DeepMind