Google · 2026 · 02 · 19 · Model · ~2 min read
Google launched Gemini 3.1 Pro (Deep Think mode)
What's actually new
- Deep Think mode. Spend minutes on a problem — visible reasoning traces, clear stops where the model checks its work. Aimed at maths, scientific reasoning, hard coding.
- 2-million-token context. Doubled from Gemini 3's million-token window for Pro users.
- Bigger jumps on tests requiring sustained reasoning — research-grade maths, multi-step scientific problems.
If you want more
Worth knowing
- Deep Think isn't always better. For everyday chat it's slow and pricey without much accuracy gain.
- Two-million-token claim degrades. Past around 800K tokens the accuracy drops noticeably — same pattern as every long-context model.
Who should care
Researchers and engineers tackling problems where wrong-but-confident answers cost real money or research credibility. Educators teaching how AI reasoning works (Deep Think makes the trace visible). Anyone whose work involves analysing huge documents — the 2M context is the underrated upgrade.
What to do about it
Use Deep Think for the kind of hard problem you'd otherwise spend a working day on. Watch the trace — it teaches you something about how the model reaches its answer. For everyday use, stick with regular Gemini 3.
Honest take
Gemini 3.1 Pro was the moment Google added what OpenAI's o-series had pioneered — visible, multi-minute reasoning — but with a longer context window than anyone else. The Deep Think mode was less a feature and more a commitment: Google decided that sustained reasoning was worth the latency, and bet it would unlock real scientific and engineering use cases. So far that bet looks right for narrow domains, less obviously useful for everyday users.
Sources
Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com