What's this about?
Two outlets reported that Google's next flagship model, Gemini 3.5 Pro, is months behind schedule. Per The Information's sources, Google refreshed the training data late last month specifically to improve coding skills — where it lags rivals most — but the results were "disappointing." Bloomberg confirms the launch has slipped past internal goals.
What exactly does "behind schedule" mean for a model?
Frontier models are trained in giant multi-month runs costing hundreds of millions of dollars. If the finished run doesn't beat the competition, a lab can't just patch it — it has to change the data or method and effectively re-run, burning months. Insiders call this the "disappointing giant model trap": Meta hit it with Llama 4 and xAI after Grok 4. Ethan Mollick notes only OpenAI has publicly stumbled through it (GPT-4.5/Orion) without losing its lead.
What's the community saying?
- Timing is brutal — the story landed the same day an open Chinese model claimed #1 spots over Fable 5; Google now looks squeezed from above and below.
- Coding is the battleground — the detail that the data refresh specifically targeted coding confirms where labs think the competitive frontier is.
- dotey's practitioner rant (separate viral post): Gemini's web app has barely improved in a year and still lacks Skills, while Codex ships multiple updates a day.