The news hit quietly on Monday: Google has reportedly shelved Gemini 3.5 Pro, the flagship model Sundar Pichai promised developers back in May. The claim comes from an institutional note by SemiAnalysis, which argues the company has moved on and is now wrapping itself in Gemini 4 hype to cover the gap. I read a lot of AI news in this job, and I have to be honest: this is the least surprising cancellation I have seen all year.
Let me take you back to Google I/O in May, because that is where the story really starts. Pichai stood on stage, told a room full of developers that Gemini 3.5 Pro was coming, and asked for patience. It would arrive next month, he said. Multiple reports say the line drew an audible groan. A room full of developers groaned at Google's own keynote. That was the first warning sign.
The Deadline Google Set for Itself
June came and went with no Pro model. Instead, Google spent July retraining the model on updated data to fix coding performance, and when that work fell short of internal expectations, it shipped Gemini 3.6 Flash on July 21 as a stopgap, alongside Flash-Lite and a cybersecurity-focused Flash Cyber variant. The stopgap was efficient, sure. It used around 17 percent fewer output tokens than its predecessor and beat it on benchmarks like DeepSWE, OSWorld-Verified, and MLE-Bench.
But here is what efficiency does not buy you: intelligence. Independent testing from Artificial Analysis gave Gemini 3.6 Flash the exact same Intelligence Index score as Gemini 3.5 Flash. A lateral shift, not an upgrade. The stopgap did not move the needle on raw capability, and the competition noticed. OpenAI and Meta staffers were reportedly mocking Google on X the same day the models dropped.
Here is the timeline in one breath:
- May: Announced at I/O, promised for the following month.
- June: No flagship arrives.
- July: Retraining falls short; the stopgap 3.6 Flash ships as a lateral move.
- August: SemiAnalysis reports the Pro model has been quietly shelved.
Eighth Place Is a Weird Look for a $4 Trillion Company
The SemiAnalysis note is blunt about where this leaves Google, and it is a genuinely strange place for this company. Gemini currently sits eighth in the Artificial Analysis Intelligence Index, trailing Muse Spark 1.2, Grok 4.5, and a top tier of Chinese open-source labs that now outperform Gemini 3.6 Flash on a consistent basis. Google broke through a $4 trillion valuation partly on the strength of Gemini 3's reception late last year. Eight months later it sits outside the top five.
There were signs of life, of course. Gemini 3.5 Pro surfaced on Arena a handful of times over the past few weeks, briefly appearing in blind testing pools before being pulled, in one case after less than an hour. Each sighting set off a fresh round of speculation on X that a launch was imminent. But the reaction to the outputs never matched the hype. Testers who ran their usual prompts against the model came back with the same verdict: incremental, not generational. One widely shared thread described the results as a lazy skeleton dressed up to look finished.
The Gemini 4 Pivot Is Hype, Not Strategy
Then came the detail that reads, to me, like the real confession. Last month, alongside the Gemini 3.6 Flash launch, Google said it had already begun its most ambitious pre-training run yet, this one for Gemini 4. Frontier pre-training runs take months of compute and iteration before a model is even ready for internal testing. That means Google was running Gemini 4 pre-training while Gemini 3.5 Pro was supposedly weeks from release. You do not start the next baby while the current one is still in the delivery room. Nobody who watches this industry would read that timeline any other way: the team moved on.
And here is where I break with the people who think Gemini 4 fixes all of this. SemiAnalysis is not optimistic, and neither am I. The structural issues that held back Gemini 3.5 Pro, particularly around coding and agentic reliability, are not the kind of problems a bigger pre-training run automatically resolves. More compute compounds what you already do well. It does not teach you to do the thing you keep failing at. Google has been unusually open about lagging behind on agentic coding, going so far as to form a dedicated DeepMind coding team to close the gap. You form a dedicated team when you admit the base organization could not handle it.
Looking back, the warning signs were everywhere:
- It kept getting pulled from Arena blind tests, sometimes in under an hour.
- The stopgap replacement scored identically to the model it replaced.
- Gemini 4 pre-training started before 3.5 Pro ever shipped.
- The company's own timeline kept moving, and nobody apologized.
Here is my take, and I know it is a hot one: Google is not failing at AI because it lacks talent, compute, or money. It is failing because it keeps announcing models before they are real, then spends the next quarter pretending the delay was always the plan. The Gemini 3.5 Pro cancellation is not a strategy. It is a confession, delivered in the one language this company still speaks fluently: a pivot to something bigger, something always just a few months away. I hope Gemini 4 is everything Google says it will be. I also remember what Google said about 3.5 Pro, and I will believe it when I can type against the model myself. Until then, eighth place is the honest ranking.
Comments