Gemini 3.5 Pro slips again. Buy models on cost-to-complete

Google's flagship model is late again. As of the second week of July 2026, Gemini 3.5 Pro is still stuck in a limited Vertex AI preview with no confirmed general-availability date — after Google promised it "next month" at I/O on May 19 and then let June slide by. The interesting part isn't the slip. It's why Google says it's holding the model back: token efficiency. When the company sitting on the most compute on earth won't ship its flagship because it burns too many tokens per task, that tells you exactly how to buy AI for your own business.

What actually happened

Per Business Insider reporting, Google cited three linked reasons for keeping Gemini 3.5 Pro in preview: token-efficiency concerns flagged by early testers, coding performance below flagship standard, and long-horizon, multi-step reasoning that fell short of the bar Google set at I/O. Google framed it as a quality decision, carrying over a lesson from Gemini 3.5 Flash, where some users "burned through tokens faster than expected."

That token line is the whole story. A model that needs more tokens to reach the same answer is simply more expensive to run at scale — even if it aces the leaderboards. Reports point to a mid-July target for release, but Google hasn't confirmed one, and the model remains gated to a small set of preview customers and testers on LMArena and Google's Antigravity platform.

Why it matters for your business

Most teams still pick a model the way they'd read a spec sheet: highest benchmark, biggest context window, newest name. That's how you end up with an agent that scores beautifully and then eats your quarterly AI budget in three weeks — which is exactly what happened to a lot of companies in Q2 2026. The metric that actually hits your card is cost-to-complete: how many tokens (and dollars) it takes to finish one real task on your workload, not a benchmark's.

Google is telling you this in the loudest way possible — by refusing to ship until the number is right. You don't have that luxury of waiting, but you have the same test available. Before you standardize on any model, run your actual jobs — the support triage, the product-description generator, the code review — through two or three candidates and measure tokens-per-completed-task and error rate, not vibes. Cheaper-per-token often loses on cost-to-complete because it takes more turns to get there. The winner is workload-specific, which is the whole point.

Sources: MarketScale (citing Business Insider).