Gemini 4 plans put larger AI models at center stage again

Gemini 4 is already shaping Google’s AI roadmap as Sundar Pichai says the company is training a larger frontier model.

The comments came during Alphabet’s second-quarter 2026 earnings call, where Pichai framed the next model generation as central to Google’s effort to stay competitive at the top end of artificial intelligence. The company is also testing Gemini 3.5 Pro with partners before a launch described as coming “as soon as it’s ready.”

Alphabet call puts Gemini 4 first

Google has not announced a public release date for Gemini 4, but Pichai said work is already underway on what the company calls its “most ambitious pre-training run yet.” That phrasing points to a larger and more expensive model-building cycle, with compute supply becoming a direct strategic constraint.

Pichai told analysts that the frontier AI market is changing quickly and that Google sees both strengths and gaps in its current models. He singled out coding and agentic coding as areas where the company wants to improve, saying internal teams are focused on that work.

The chief executive’s message was designed to answer a direct investor concern: whether Gemini can remain close to the best-performing AI systems. Pichai said Google is “both very committed and very confident of being at the frontier.”

Larger models raise TPU pressure

The clearest operational signal came in Pichai’s comments on Tensor Processing Units, Google’s in-house AI chips. Asked how the company balances internal use with outside customer demand, he said the first priority is allocating enough TPU capacity to compete in AGI development.

That priority matters because frontier model training depends on large clusters of specialized compute. If more TPUs are reserved for Google’s own model work, external cloud customers may face a tighter supply environment, depending on how much new capacity Alphabet brings online.

Pichai also linked Gemini 4 directly to model scale. He said the next generation of frontier systems will need “much larger base models” and added that Google is “being very ambitious with it.”

The statement does not reveal parameter counts, training budgets, benchmark targets or release timing. It does show that Google sees scale, coding performance and agent-style tasks as the main battlegrounds for the next Gemini cycle.

Flash cadence keeps pressure on rivals

Gemini 4 is not the only part of the roadmap. Pichai also pointed to continued work on Gemini 3.x Flash models, with more progress expected in agentic coding and related capabilities.

Google is targeting releases “almost at a monthly cadence,” according to Pichai. That faster rhythm gives the company a way to ship incremental improvements while the larger Gemini 4 training run continues in the background.

The split strategy is familiar in AI product development: smaller and faster models can serve cost-sensitive uses, while larger frontier systems compete for headline performance. For Google, that means Gemini Flash can keep developers and enterprise users engaged even before the next flagship model arrives.

Three paths for Google’s AI race

If Gemini 4 delivers meaningful gains in coding and agentic work, Google could strengthen its position in enterprise AI and developer tools. The macro effect would be more pressure on companies to raise AI infrastructure spending, while Google would gain a stronger sales case for Gemini and its cloud platform.

If the model is delayed or falls short of frontier expectations, the impact would move in the opposite direction. Google would still have Flash updates to sustain product momentum, but the wider AI sector could read the delay as evidence that larger training runs are becoming harder to turn into visible performance gains.

A third path sits between those outcomes. If Gemini 4 improves steadily but not dramatically, Google may lean more heavily on monthly Flash releases, TPU allocation discipline and integration across its products. The key open questions are how much compute Google can spare for outside customers, when Gemini 4 is released, and whether coding performance becomes a visible advantage rather than a stated priority.