Gemini 3.5 Pro Is Late Again, and Your Roadmap Is the Collateral

As of mid-August 2026, Gemini 3.5 Pro was still unreleased, months past a deadline originally set nearer the start of the summer, and Forbes columnist John Werner counts it as the third date Google has blown on the model ^[1][2][3]. What Google shipped in the meantime was 3.6 Flash, a cheaper workhorse, which means the product naming ladder has now stepped past the model that teams are actually waiting on ^[4][11].

The substitution is worth reading carefully, because it is not a like-for-like consolation. Per the outlet Andrew.ooo, cited by Werner, Google said it needed more time on coding and reliability, and rather than rush Pro it shipped 3.6 Flash with improved coding, knowledge and multimodal performance at lower cost, using up to 65% fewer output tokens than 3.5 Flash ^[4][5]. If your binding constraint was cost or latency per call, that is a real gift. If your binding constraint was reasoning headroom, and your feature only works at the Pro tier, you have received nothing and lost a quarter.

The explanations on offer are numerous, unverified, and mutually incompatible, which is itself the signal. Andrew.ooo also points to senior-researcher churn ^[5]. Werner writes that internal Google reports have reps saying 3.5 Pro is "being tested with partners," in the manner of Anthropic's Project Glasswing or OpenAI's pre-launch holds, while noting that Google itself has not made the too-powerful-to-release argument ^[6][7][8]. A separate theory is engineering attention: 9to5Google's Abner Li reported on July 9 that Gemini lead Josh Woodward had solicited user fix requests and published a top-ten list including Workspace integration reliability, deep research improvements, watermark removal from Nano Banana, and mobile scrolling bugs ^[9][10]. Werner also raises an internal debate over how much code AI agents should write ^[12]. And even the surface facts are contested: Geeky Gadget and Andrew.ooo disagree on whether Google's own Gemini page still lists 3.5 Pro as "coming soon" ^[13].

The most consequential version, if true, is the one that has nothing to do with polish. Tech Times' Eloise Jones writes that Google reportedly chose to restart pre-training, and that because pre-training sets a model's capability ceiling while fine-tuning and RLHF only refine within it, a restart concedes a structural problem rather than a finishing problem ^[14][15]. Read as schedule risk, that reprices the wait: you are not waiting on a release-candidate cycle, you are waiting on the longest and most expensive phase of model construction, started over ^[16].

Around this sits the market's read. Wired reports that Jeff Dean and other prominent researchers have left for OpenAI and Anthropic, and Alphabet stock is down roughly 12% since January ^[17][18].

For operators the lesson is narrow and cheap to apply. A named future model is a marketing artifact, not a dependency you can plan against, and no vendor owes you the reason for a slip. Watch for whether the "coming soon" language on Google's page firms up or vanishes ^[13], whether a later Pro or a 4-series release jumps the queue and quietly makes 3.5 Pro vestigial ^[11], and whether Google commits to any date at all this time. Then audit your own gates for the version numbers you have written into them.