The company that gave us "move fast and break things" just chose slow and maybe-broken instead.
The Summary
- Google has delayed Gemini 3.5 Pro by months, missing internal targets because the model isn't good enough at coding
- This marks a shift from Google's historic ship-and-iterate culture to holding releases until quality bars are met
- The coding weakness matters because code generation is the killer app proving AI agents can actually ship value
The Signal
Google is doing something it almost never does: admitting a product isn't ready and actually waiting to ship it. Gemini 3.5 Pro was supposed to be out by now. Instead, it's sitting in Google's labs because it can't code well enough to meet whatever bar the company set internally.
The coding piece is the tell. Every frontier model lab right now is in an arms race over code generation. Not because developers are the only market, but because code is the proving ground for whether these models can do real work. If an AI can't reliably write a function, debug an API, or refactor a messy script, it's not going to manage your calendar or negotiate your vendor contracts.
"Code generation is the killer app proving AI agents can actually ship value."
Google's delay suggests one of three things:
- Gemini 3.5 Pro genuinely can't compete with GPT-4 or Claude on coding benchmarks
- Google's internal quality bar just got higher, possibly because enterprise customers are pushing back on hallucinations and broken outputs
- The company is seeing what happens when you ship half-baked AI tools and decided to pump the brakes
The timing matters. OpenAI just launched GPT-5 with substantially better reasoning. Anthropic's Claude is the model developers actually want to use for code. Google is already playing catch-up in a race where second place means your models become commoditized infrastructure instead of defensible moats.
This isn't just about one delayed model. It's about whether Google can ship AI products at the pace the market is moving. The company has more AI research talent than anyone. It invented the transformer architecture that powers every major LLM. But research excellence doesn't mean product velocity, and right now the game is who can ship useful agents fastest.
The Implication
Watch whether Google's competitors pounce on this delay window. If OpenAI or Anthropic can lock in enterprise coding customers now, those contracts don't flip easily. The pattern that's emerging: the AI lab that ships usable products wins, not the one with the best papers. Google pioneered the tech. That matters less every month.