Google showed off Gemini 3.5 Pro at its I/O developer conference on May 19 and told the room it was targeting general availability in June. It is now the second week of July, and the model still is not out. It sits in a limited Vertex AI enterprise preview — open to a handful of customers through Google’s developer platform — with a launch now reportedly aimed at July 17, a date Google itself has declined to confirm.
The company has been unusually specific about why. Early testers flagged three linked problems: token efficiency that ran too rich, coding performance short of a flagship tier, and long-horizon, multi-step reasoning that missed the bar Google set on stage in May. Reporting on the latest slip says Google scrapped the base model for a ground-up rebuild and is running extra pre-training to close the gaps, with math and other reasoning tasks singled out.
Of the three, token efficiency is the tell. In 2026, “intelligence per dollar” stopped being a marketing line and became a line item. Enterprise buyers now compare models on the cost to complete a task, not raw benchmark scores, and vendors have started publishing average token usage per task right on their model cards. A model that answers correctly but burns more tokens getting there loses the deal. Google holding back a finished-looking model to fix its cost curve is a direct read on where the market’s leverage now sits.
Our take: A blown launch window used to be an engineering footnote. This one is a strategy signal. Google could have shipped Gemini 3.5 Pro in June and let the benchmarks talk — and chose instead to hold it until it wins on the metric buyers actually expense. That is a quiet concession that the frontier is no longer scored on who is smartest, but on who is smartest per dollar. The risk is the calendar: every week in preview is a week rivals sell into. OpenAI’s GPT-5.6 is benchmark-topping but gated to a short list, and Anthropic’s Fable 5 now bills by the token. Google’s opening is real — a broadly available good model beats a locked-down great one — but only if “broadly available” actually arrives.
The backdrop makes the delay louder. On June 22, Alphabet shares fell about 5% — an estimated $225 billion in market value — after a run of high-profile researcher departures, including a Gemini co-lead who co-wrote the 2017 transformer paper that underpins nearly every large language model in use today. Wall Street split on the read: Morgan Stanley, JPMorgan and Goldman Sachs kept bullish ratings, citing Google’s compute, Cloud and Workspace distribution, while the sell-off priced in a quieter worry — whether the roadmap, and the people driving it, are still out front.
What to watch
- July 17. It is a reported target, not a signed launch post. Slip it again and the story stops being “careful” and starts being “stuck.”
- The token-per-task number. When Gemini 3.5 Pro ships, the figure that matters is not its benchmark score — it is cost-to-complete against the cheap open models enterprises are already routing work to.
- Availability, not access. Google’s whole edge here is shipping broadly while rivals gate. A staggered or waitlisted rollout would throw away the one advantage the delay is supposed to buy.