Google launches Gemini 3.1 Ultra with 2M-token context

2 million tokens. I had to read that twice. That's roughly 1,500 average codebases shoved into a single context window — not chunked, not summarized, just... there.

Here's the thing: this makes our entire RAG setup for document-heavy clients look like unnecessary overengineering. We've spent real hours on chunking strategies and embedding pipelines. With 2M tokens you can just send the whole thing. That's not a small deal.

The native multimodal piece is what's getting buried under the context headline. No transcription step, no image-to-text preprocessing — it reasons across images, audio, and text natively. That quietly kills a whole class of pipelines we've been bolting together.

Honestly? I'm skeptical of the latency and cost story at 2M tokens. Google hasn't published per-token pricing yet (classic), and inference at that scale is never free. Don't architect anything around this until you've run real benchmarks.

Sandboxed code execution is them going hard at the agentic use case. We're going to test it against our current workflow this week — if it holds up, some of what we've built in the last six months is getting simplified.