Google DeepMind has pushed back the launch of Gemini 3.5 Pro to July 17, 2026, after scrapping the existing Gemini 2.5 Pro architecture entirely in favour of a ground-up rebuild. The delay follows a missed June launch window publicly promised by CEO Sundar Pichai at Google I/O, and arrives against a backdrop of significant talent departures that have rattled investors and raised questions about Google’s competitive position at the AI frontier.
What Happened
Google confirmed that Gemini 3.5 Pro would not ship until July 17 after internal testing identified three performance gaps the existing Gemini 2.5 Pro architecture reportedly could not close: mathematical reasoning, scalable vector graphics (SVG) scene generation, and overall image quality. Rather than patch the existing model, the company ran a new pre-training cycle from scratch — an unusual and costly decision that reflects the scale of the competitive pressure Google is under.
The rebuilt model introduces a two-million-token context window, a Deep Think Reasoning Layer for complex multi-step problem-solving, and autonomous workflow capabilities designed to compete with Anthropic’s Fable 5 and OpenAI’s GPT-5.6 series. Google gathered feedback through its Antigravity testing platform and LMArena, incorporating lessons from the Gemini 3.5 Flash release into the full rebuild.
Why It Matters
The delay is more than a calendar slip. It coincides with an unprecedented wave of senior departures from Google DeepMind. In a single week, four prominent researchers left the organisation: Noam Shazeer — a Gemini co-lead and co-author of the landmark 2017 paper “Attention Is All You Need,” which introduced the transformer architecture underpinning virtually all modern large language models — departed for OpenAI. Nobel laureate John Jumper, along with researchers Jonas Adler and Alexander Pritzel, moved to Anthropic.
Markets reacted sharply to the combined news. Alphabet shares fell roughly five percent in a single trading session in late June, erasing approximately $225 billion in market capitalisation. Investors have grown increasingly concerned about whether Google can retain the research talent necessary to compete at the frontier of large language model development over the long term, particularly as rivals continue to ship capable models on tighter timelines.
Background and Context
Gemini 3.5 Pro is Google’s most significant model release of 2026, intended to reclaim benchmark leadership ceded to rivals in recent months. The competitive landscape has shifted materially: Chinese AI models now account for more than 30% of weekly token usage at US companies, peaking at 46% in some weeks, according to OpenRouter data. Models like GLM-5.2 are landing within a percentage point of Anthropic’s Opus 4.8 on key agentic benchmarks at roughly one-fifth of the cost — a dynamic that is pricing US frontier labs out of middle-tier enterprise workflows and accelerating the shift to lower-cost alternatives.
Within Google, the Gemini programme has faced recurring scrutiny over the gap between benchmark claims and real-world performance. The decision to rebuild from scratch rather than release an incrementally improved model suggests that leadership recognises the need to deliver a model that can genuinely move the needle — not merely improve marginally on its predecessor. The Antigravity platform and LMArena data gave the team a clearer picture of where the architecture was failing, making the case for a full pre-training cycle rather than targeted fine-tuning.
What Comes Next
Google now targets July 17 as a firm general availability date for Gemini 3.5 Pro. Given the model’s track record of slippage, enterprise developers and benchmark watchers will be looking for confirmed access before updating deployment plans. A successful launch could help stabilise Alphabet’s share price, restore confidence in Google’s research capability, and give developers a domestic frontier option to weigh against the growing appeal of lower-cost Chinese alternatives.
Longer term, the talent departures raise structural questions that a single model release cannot fully answer. Sustaining frontier AI research requires institutional depth and the ability to retain researchers who have options at every major lab. Separately, the ongoing development of a voluntary AI model review framework being negotiated with leading US labs means that how frontier models perform on standardised capability and safety benchmarks is increasingly becoming a matter of public record — adding another dimension to the pressure Google faces heading into the Gemini 3.5 Pro launch.
