Pichai admits Google trails in AI coding as Gemini 4 enters pretraining

Alphabet CEO Sundar Pichai confirmed on the company's Q2 2026 earnings call that Gemini 4 has entered pretraining, calling it "our most ambitious pretraining run yet." The disclosure came alongside a pointed admission: Google needs a larger Gemini 4 base model to compete "at the frontier level of where the frontier will be when Gemini 4 comes out" — language that acknowledges Google is training toward a moving target its rivals are actively pushing forward.

The gap Pichai named directly

Pichai specifically flagged coding and agentic coding as areas requiring improvement, an unusually direct admission from a CEO typically measured in how he frames competitive gaps. The comment lands against a backdrop of two high-profile departures: Gemini co-lead Noam Shazeer and AlphaFold researcher John Jumper both left for OpenAI and Anthropic in June 2026, reportedly amid internal concern about Google's AI coding capabilities relative to competitors.

The coding gap has a concrete, visible symptom. Gemini 3.5 Pro — announced at Google I/O in May 2026 with a promised June release — still has no confirmed launch date, delayed specifically by coding performance issues. In its place, Google shipped Gemini 3.6 Flash on July 21, an update to the smaller Flash tier rather than a flagship release: 17% fewer output tokens than 3.5 Flash, and a jump to 49% on the DeepSWE coding benchmark, up from 37%. That's a real improvement, but it's happening in the tier below the one Google actually needs to ship to compete for enterprise and developer coding workloads.

Why the timing matters

Pichai's admission comes the same week Alphabet reported Google Cloud revenue up 82% to $24.8 billion, driven substantially by AI infrastructure demand — meaning Google's cloud business is currently thriving on AI demand even while its own model lineup lags on one of the highest-value use cases (coding and agentic tasks) driving that demand elsewhere in the industry. Anthropic's Claude and OpenAI's GPT-5-class models have built substantial developer mindshare specifically around coding performance, and Google shipping a delayed Pro-tier model and pretraining a not-yet-named Gemini 4 puts a concrete, if unconfirmed, timeline gap between Google and its two closest frontier competitors.

What comes next

No release date has been set for either Gemini 3.5 Pro or Gemini 4. Pichai's framing — training Gemini 4 to compete at wherever the frontier will be by the time it ships — implicitly concedes that Google is training against a target that keeps moving while the model is still in pretraining, rather than against today's competitive landscape. For a company whose cloud revenue is increasingly tied to being seen as an AI leader, the gap between infrastructure momentum and model-tier competitiveness is now the more consequential story than the earnings beat itself.