Google Is Making a Big Bet
Tomorrow, July 17, Google DeepMind's flagship model, Gemini 3.5 Pro, finally drops.
It was supposed to land in June. Instead it got pushed back over a month. And this wasn't a tuning issue — Google made a brutal call: scrap the entire Gemini 2.5 Pro base model and retrain from scratch.
Not a paint job. A teardown and rebuild. That costs hundreds of millions of dollars and months of GPU time.
Why Go Through With It
Three ceilings got hit: math reasoning, SVG scene generation precision, and overall image quality.
Frontend code especially. Old Gemini had been getting cooked by Anthropic's Fable 5 — redundant UI code, broken layouts, rough visuals. The retrain was aimed straight at that gap.
From leaked tests, it landed. In anonymous LM Arena runs, Gemini 3.5 Pro "mogged" Fable 5 on frontend and visual code tasks — that's the word developers started using. Total domination.
But hold the hype. On the hardest agentic tasks, repo-level engineering, and long-horizon work, it still loses to Fable 5 and GPT-5.6. Strong at frontend doesn't mean strong everywhere.
Two Million Tokens — Impressive, but Unverified
The new model claims a 2-million-token context window.
For scale: double Gemini 3.5 Flash, roughly 8x Fable 5 (256K), and 15x+ the standard GPT-5 (128K). In theory, you could feed it an entire enterprise codebase.
But here's the catch — Transformer attention scales quadratically with sequence length. Doubling context doesn't double cost; it jumps by orders of magnitude. And researchers have long documented the "lost in the middle" problem: stuff placed in the middle of giant contexts tends to get ignored.
So is 2 million a real capability or a marketing number? Won't know until independent long-context retrieval benchmarks land.
The Talent Drain Is the Real Story
Bigger than the model itself: the people problem at Google's AI team.
On June 18, Noam Shazeer — co-author of "Attention Is All You Need" and co-lead of the Gemini team — announced he was leaving for OpenAI. The next day, Nobel laureate John Jumper left DeepMind for Anthropic.
Two core researchers gone, and on June 22, Alphabet's market cap dropped $225 billion in a single session — a 5% slide.
Models can be retrained. People are harder to replace. That's what Google should actually worry about.
July 17 Is a Crowded Day
And Google isn't the only one showing up.
- Google Gemini 3.5 Pro launches
- DeepSeek V4 stable goes GA the same day
- xAI Grok 4.5 opens public beta
Three on one day, an obvious head-to-head. Pricing for Gemini 3.5 Pro lands around $15 input / $60 output per million tokens — 10x Flash, but notably cheaper than GPT-5.6.
Google's also building Nano Banana Pro on the same base, targeting OpenAI's GPT-Image 2. One pretrain, two product lines. Smart math.
So we'll see tomorrow. This is Google's "quality-first" gamble, and the frontend code turnaround is real. Whether it can crack the moats Fable 5 and GPT-5.6 have built — that's a question only real-world usage will answer. We wait for benchmarks.