Google Delays Gemini 3.5 Pro Over Agent Token Costs

Google missed its June 30 target for Gemini 3.5 Pro after testers flagged excessive token consumption in long agentic tasks, per BuildFastWithAI. Not a benchmark miss. A cost one: multi-step agent runs were burning through token budgets before finishing the job. Google delayed rather than ship a model builders couldn't afford to run.

A Polymarket market tracking a by-June-30 release had already priced this as close to certain. It settled No the same day the deadline passed.

The bigger point: the hard problem for frontier labs isn't whether a model can complete a long task anymore. It's whether it can do that without burning through the budget and losing the thread. That distinction matters for anyone building agent products. A model that reasons well but runs up the bill before finishing doesn't ship, no matter what the benchmarks say.

Google chose cost over calendar. The real test is whether Gemini 3.5 Pro ships with a real fix for agentic token burn, or slips again. The deadline passed June 30, and the Polymarket market settled the same day.