Google DeepMind's Gemini 3.5 Pro is reported to launch on July 17 after a six-week delay in which Google scrapped its original base model entirely and restarted pretraining, after engineers reportedly found structural failures in recursive tool-calling, the kind of multi-step, model-calls-a-tool-which-calls-the-model pattern that modern agent workloads depend on. That is a much bigger admission than a normal launch slip: it means the earlier pretraining run itself was judged unfit rather than just needing more fine-tuning or patching, and Google chose to redo the expensive part rather than ship something with a known structural weakness in agentic use. Leaked specifications, not yet officially confirmed by Google in full at launch, point to a 2-million-token context window, double Gemini 2.5 Pro's 1 million cap, a "Deep Think" extended-reasoning mode gated to the roughly $250 per month Ultra tier, and API pricing near $1.25 per million input tokens and $10 per million output tokens. For developers, a 2-million-token window is meaningful mainly for three use cases: reasoning over an entire codebase or monorepo in one pass instead of chunking it, retrieval-augmented generation over large document corpora without aggressive splitting, and long-horizon agent memory that does not need external summarization to stay within context. The number itself proves nothing on its own, though: the recall and reasoning quality has to hold across the full length of the window rather than degrading past some fraction of it, which has been the norm across the industry's previous long-context claims, so the useful verification is independent long-context benchmarks run after launch, not the headline figure. The launch also lands the same week Moonshot released the 2.8-trillion-parameter open-weight Kimi K3 with a 1-million-token window, sharpening the question of what a closed model at $1.25 input has to offer against a comparably capable open alternative.