Google Confirms Gemini 4 Training Underway, Says Next-Gen Model Needs a Larger Base Model to Compete at the Frontier

Google has confirmed that Gemini 4 is currently in training, describing the effort as the company's "most ambitious pre-training run yet," according to statements made around the launch of Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber this week.

The confirmation comes as Gemini 3.5 Pro — originally expected in June — remains in partner testing, with Google saying it will ship "as soon as it's ready." Work on Gemini 4 is proceeding in parallel, positioned by Google as the company's next attempt to hold the frontier position in large language models.

Pichai on scale and confidence

During Alphabet's Q2 2026 earnings call, CEO Sundar Pichai fielded analyst questions about whether Gemini models can stay competitive at the frontier. His response did not include benchmark numbers or a parameter count, but it did confirm the core technical direction:

"For the next generation of frontier, you're going to need much larger base models. We are now training Gemini 4, and we're being very ambitious with it."

Pichai added that he is "very excited by the progress" seen internally and said he expects Google to compete "at the frontier level of where the frontier will be when Gemini 4 comes out" — an acknowledgment that competitors including OpenAI, Anthropic, and xAI will likely have shipped their own next-generation models before Gemini 4 arrives.

Asked separately about how Google allocates its Tensor Processing Units (TPUs) between internal AGI research and external cloud customers, Pichai said the "first priority is making sure we are allocating what we need to compete at the frontier in terms of AGI development," calling that the "foundation for everything we do."

What's still missing

Google has disclosed no specifics on Gemini 4: no context window size, no pricing, no benchmark scores, and no confirmed parameter count. Pichai did flag coding and agentic coding as current weak points relative to competitors, saying teams are "very, very focused" on closing that gap.

On the smaller end of the lineup, Google said it intends to keep iterating on Gemini 3.x Flash models on close to a monthly cadence, with continued focus on agentic coding capability.

Based on the release timing of Gemini 1.0, 1.5, 2.0, and 3.0, Gemini 4 would be expected to launch in November or December 2026, though Google has not committed to a date.

What this means

This is a teaser, not a launch — Google has confirmed a training run and a strategic rationale (bigger base model, more compute) but has disclosed zero verifiable specs. The comments read as an attempt to reassure investors and developers that Google isn't ceding frontier-model ground while Gemini 3.5 Pro's delay lingers. The emphasis on needing a "much larger base model" suggests Google is betting on scale over efficiency gains for this next jump, a departure from the industry's recent lean toward smaller, more efficient models with better post-training. Until Google publishes actual benchmarks, pricing, or a context window figure, everything about Gemini 4's real-world performance remains speculative.