A June Roadmap Built on a July Model
A product lead in Berlin had penciled Gemini 3.5 Pro into a June release plan, with a customer demo booked around it. Then Google moved the model to 17 July 2026 and, more unusually, said it had scrapped the base architecture underneath it and rebuilt from scratch. The demo now rests on a model that does not exist yet.
This is the second decision hiding inside one announcement. Google made a hard call about its own product. The Berlin team made a softer one months earlier, when it treated a promised launch date as a fixed input. Only one of those two decisions was disciplined, and it was not the buyer's.
Why Google Threw Away a Finished Model
Google DeepMind did not tune the old model; it abandoned it. Leaked internal evaluations showed the previous Gemini 2.5 Pro base architecture held up on plain text but broke down under recursive tool-calling and complex, multi-layer layouts, and stumbled on mathematical reasoning. Rather than ship a flagship that failed where its rivals are strong, Google rebuilt the base and reset the date.
The replacement targets a two-million-token context window, a deeper reasoning layer, and steadier long-horizon workflows, aimed squarely at OpenAI's GPT-5.6 and Anthropic's Fable 5. The backdrop was not calm: senior researchers left for OpenAI and Anthropic in June, and Alphabet shed roughly 225 billion dollars of market value in a single session on 22 June. Shipping a weak flagship into that would have been the easy, expensive mistake.
The Discipline Most Teams Skip
The hard part of a sunk cost is that it feels like progress. A build that is ninety percent done, with real money and months already spent, is exactly the project teams talk themselves into shipping. The rational test ignores what a project cost and asks only what finishing it is worth from here. Google applied that test to a near-complete model and killed it. Most organizations, faced with the same nearly-done build, ship it and manage the fallout.
The lesson is not that delays are good; it is that the decision to stop should turn on future value, never on how much is already invested. If a project cannot clear the bar it was built to clear, the sunk months are gone whether you ship or not, and shipping only adds the cost of the cleanup. Google just paid that lesson forward in public, at flagship scale.
What This Changes for Your Own Roadmap
Treat every unreleased model as a dated assumption, not a commitment. If a launch sits on your critical path, write down the date you are counting on, the fallback model you will use if it slips, and the last day you can switch without breaking a customer promise. A dependency you have written down is one you can manage; one that lives only in a slide is one that manages you.
The firms that got hurt this quarter were the ones that signed deliverables against a launch that had not happened. Availability is a procurement variable, and this month it moved for everyone at once. Plan against what has shipped, keep a tested alternative one switch away, and let the vendors absorb their own slipped dates instead of passing the risk to your customers.
Read next: Why Is the Most Valuable AI Lab No Longer the Most Famous One? | When the Most Invested Believer in AI Steps Back, What Are They Telling You?
