Google Begins 'Most Ambitious' Gemini 4 Pretraining as 3.5 Pro Remains Stalled

Google is asking the AI industry to look past its troubled flagship. On July 21, the company confirmed it has begun what it calls its "most ambitious pretraining run yet" for Gemini 4 — the next-generation model that will succeed the repeatedly delayed Gemini 3.5 Pro — while shipping three new Flash-tier models as a stopgap.

The Announcement

In a blog post accompanying the release of Gemini 3.6 Flash, 3.5 Flash-Lite, and the security-tuned 3.5 Flash Cyber, Google dropped the Gemini 4 news almost as a footnote. "We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress," wrote Logan Kilpatrick, Google's AI lead, on X. The company offered no timeline, no benchmark previews, and no commitment on when training might complete.

The subtext is impossible to miss. Gemini 3.5 Pro, the flagship that was supposed to compete with GPT-5.6 and Claude at the frontier, has now missed multiple launch targets. Google scrapped the original 3.5 Pro base model and restarted pretraining once already, and the second attempt is still not ready to ship. Rather than continue promising a model it cannot deliver, Google is asking the market to bet on the generation after it.

Why It Matters

The move is both pragmatic and risky. Google has the compute infrastructure to pull off a massive training run — including its new Frozen v2 chip, which claims 6 to 10 times the efficiency of its predecessor TPU — and a clean architectural reset could leapfrog competitors rather than chase them. But pretraining a frontier model takes quarters, not weeks, and the competitive calendar is unforgiving. DeepSeek V4 hits stable release on July 24, Kimi K3's free weights go live July 27, and both Anthropic's Claude and OpenAI's GPT-5.6 family are shipping into the gap Google is leaving open.

The Flash-tier releases Google did ship on July 21 are genuinely competitive — Gemini 3.6 Flash uses 17 percent fewer output tokens and costs $7.50 per million output tokens, undercutting its predecessor's $9 — but these are high-volume workhorse models, not flagship reasoning systems. Enterprises making platform decisions this quarter need a frontier roadmap, not a pretraining announcement.

The Bigger Picture

Google's bet on Gemini 4 reflects a broader industry reality: the competitive center of gravity is shifting. Frontier benchmarks still win headlines, but the high-volume Flash tier — where most production work actually runs — is becoming the real battleground. Google strengthened that tier on July 21 with three models across efficiency, throughput, and security, choosing to compete on price and token economics while it rebuilds its flagship from the ground up.

Key takeaway: Google is effectively asking the market for patience it may not have. The Gemini 4 pretraining run is the company's biggest swing yet, but with open-weight models from DeepSeek and Moonshot arriving this week and rivals shipping new flagships every quarter, the window for a clean recovery is narrowing fast. The Frozen v2 chip gives Google a hardware advantage — whether it translates into a model that resets the competitive landscape depends entirely on what comes out of this training run.