Skip to content
NewsModel release

Leaks point to July 17 for Gemini 3.5 Pro after Google scraps its 2.5 Pro base

· by Pondero Newsdesk

The short version

Google DeepMind is targeting July 17 for Gemini 3.5 Pro, per multiple leak sources, after abandoning the Gemini 2.5 Pro base in favor of a full pre-training restart that pushed the model from June into July.

Leaks point to July 17 for Gemini 3.5 Pro after Google scraps its 2.5 Pro base

Gemini 3.5 Pro missed its June launch window because Google DeepMind threw out its Gemini 2.5 Pro foundation and restarted pre-training from scratch. Leak channels "Universe of AI" and "World of AI" now cite July 17 as the new target date, and TechTimes reported the same window in late June, citing a source familiar with the matter. Google has not confirmed the date.

What changed and why

Google DeepMind originally planned to build Gemini 3.5 Pro on top of the Gemini 2.5 Pro base. That plan was scrapped. Per Geeky Gadgets, the company decided the 2.5 Pro base could not close performance gaps against OpenAI's GPT-5.6 and Anthropic's models in the areas where it mattered most: mathematical reasoning and SVG scene generation. Rather than ship an incremental update, DeepMind chose a new pre-training run on a stronger foundation.

Google also cited token efficiency issues flagged by early testers and lessons from the Gemini 3.5 Flash rollout as reasons for the delay, per TechTimes citing a source familiar with the matter. A Google spokesperson declined to comment on the timeline.

The restart compounds a separate personnel challenge. In the week of June 21 to 27, four senior Gemini researchers announced they were leaving Google for Anthropic, per TechTimes. The departures follow a broader talent movement out of Google's AI division that has been running for several months.

What the model is supposed to deliver

Three capabilities appear consistently across the leak channels, per Geeky Gadgets:

A 2 million token context window, making it the largest context capacity in any production model at the time of its expected launch. Gemini 3.5 Flash, which launched at Google I/O on May 19, already topped its predecessor on coding and agentic benchmarks at lower cost. Pro is positioned to close remaining gaps on harder reasoning and long-context retrieval.

A "Deep Think" reasoning mode, aimed at multi-step logic and math, restricted to $250-a-month Ultra plan subscribers. Pricing for standard API access is expected to track prior generations at approximately $15 per million input tokens and $60 per million output tokens, with a 10x premium applied to the Deep Think mode, per TechTimes. None of these pricing figures are confirmed by Google.

Stronger performance on front-end design, SVG generation, and 3D modeling tasks. These are areas where the 2.5 Pro base reportedly underperformed and where the new pre-training is expected to close the gap against competitors.

Why the timing matters for developers

Gemini 3.5 Pro is currently the only major upcoming frontier model release without a government access restriction. GPT-5.6 launched in June under restrictions limiting it to roughly 20 government-approved partner organizations, after OpenAI's own internal testing found the model crossed a cybersecurity capability threshold that triggered White House review. Anthropic's Fable 5 was taken offline on June 12 under a Commerce Department export control directive; it was still offline for general users as of late June.

For developers who built long-context workflows around Fable 5 before June 12, Gemini 3.5 Pro's 2M token window directly addresses the capacity gap. That the model arrives without government restriction is a practical advantage in the short term, though TechTimes noted the de facto capability-gating regime is informal and case-by-case.

The question hanging over the July 17 date is whether Google actually ships on that day. The model has already slipped once from its original June target. No official launch page or API preview is live as of July 7.

What to watch next

Google has not scheduled a public announcement for July 17, and no preview API access or AI Studio listing appeared as of this writing. Watch for a Google DeepMind blog post or Gemini API changelog entry closer to that date. The second thing to track is whether Gemini 3.1 Pro on Terminal-Bench 2.1 scored 70.7% (18 points below GPT-5.6 Sol's 88.8%), and whether the 3.5 Pro closes that gap enough to face the same government scrutiny that restricted its competitors.

Sources