Claude Opus 4.6 released with 1M token context window

Okay, every RAG pipeline we've built in the last two years just got a question mark over it. Not because they're broken — but because chunk-embed-retrieve was always a workaround for context limits that no longer exist at this scale.

Here's the thing though: 1M tokens isn't free. Opus is already the expensive tier, and at that scale you're looking at real money per call. For one-off analysis? Fine. For anything running in a loop — you still need to do the math.

The agentic angle is where this actually gets interesting. Drop an entire Node.js monorepo into context and ask it to trace a bug across 40 files without retrieval tricks. That's genuinely new. We've been duct-taping multi-step agent chains together to compensate for exactly this limitation.

"Record-breaking benchmarks" — I've learned to squint at those (every model release has them). But 1M context isn't a benchmark, it's a spec. It either works or it doesn't.

Start with your biggest codebase analysis task — the one you've been splitting into pieces. Try it in a single shot. That's the honest test.