Google Ships Flash Instead of Pro While OpenAI Loses Its Two Best Operators

Google promised its top-tier Gemini 3.5 Pro in June, has not shipped it, and has offered no explanation; on the 13th it released the lightweight Gemini 3.7 Flash instead ^{[1] [2]}. That came three weeks after Gemini 3.6 Flash on the 21st of the previous month ^[3], which means the frontier leader that opened this cycle with Gemini 3.0 on its own TPUs ^[4] has spent the last month shipping small models and skipping the big one.

The scoreboard explains the reticence. Artificial Analysis gave Gemini 3.7 Flash an intelligence score of 56 and ranked it ninth among all models ^[5]. Of the seven other current major models the firm listed, six scored above it: Claude Fable 5 at 62, Grok 4.6 at 61, Kimi K3 at 60, Qwen 3.8 Max at 58, and both Muse Spark 1.2 and GPT-5.6 Tera at 57 ^{[6] [1]}. Only DeepSeek V4 Pro trailed, at 53 ^[6]. A lightweight model losing to flagships is not scandalous. A lightweight model being the only thing you have to release is.

The organisational reading is worse than the benchmark. Jeff Dean, chief scientist, left on the 5th after 27 years, having joined in 1999 as employee number 30 ^[7]; Alphabet fell 4.03% that day ^[8]. Sanjay Ghemawat, Oriol Vinyals and Quoc Le are joining him to found a public-interest organisation called Discovery Loop ^[9]. That follows Noam Shazeer's move to OpenAI and John Jumper's to Anthropic in June ^[10]. Demis Hassabis takes the chief scientist title with Koray Kavukcuoglu running DeepMind ^[11].

Titles are not the story here. According to the Financial Times, Hassabis agreed to lead only long-term research and handed immediate R&D and commercialisation authority to Sergey Brin ^[12], who stepped down as president in 2019 and returned to hands-on Gemini work in 2023 ^[13]. The FT reads the exits as a collision between London's academic culture and Mountain View's commercial direction ^[14], and Hassabis is staying in London ^[15]. Reuters reports Google decided on the 6th, the day after Dean left, to absorb DeepMind's non-technical support teams into headquarters ^[16]. Kavukcuoglu has been working out of Mountain View since last year ^[17]. Read together: the research lab is being pulled into the parent, and a founder is holding the ship-it pen.

OpenAI's strain is financial rather than technical. Its valuation has barely moved from $852bn, and it carried that same figure into the $7bn buyback of current and former employee shares on the 10th ^[18]. Buybacks are where paper becomes cash, and pricing one at a flat mark tells employees the up-and-to-the-right period has paused. Meanwhile Fidji Simo, who ran the applications division and was regarded as second-in-command, resigned last month to treat a chronic illness ^[19], and Brad Lightcap, COO since 2018, left on the 11th ^[20]. Losing the person who ran the product business and the person who ran operations inside a month is a hole in execution capacity, not in vision.

What to watch: whether Gemini 3.5 Pro ships under that name or quietly becomes a higher number; whether Brin's commercialisation authority produces a flagship within the quarter; whether OpenAI names replacements or redistributes the COO and applications roles; and whether Nvidia's push into open models, which it advanced on the 11th, gives buyers a reason not to wait on either lab ^[21].