Claude's three-model structure finally makes sense after watching this breakdown

The framing that actually landed: Haiku 4.5 for speed and bulk when depth is not what you need. Sonnet 4.6 as the everyday workhorse for writing, coding and analysis. Opus 4.6 reserved for the genuinely difficult problems where multi-layered reasoning is what produces the right answer rather than just a fast one.

The 1 million token context window on both Sonnet and Opus is the number that changes the scope of what you bring into a session. Entire codebases, long document libraries, extensive research archives. The retrieval accuracy claim at that scale is worth testing on your specific documents rather than taking on faith, but the capability itself opens workflows that were not practical before.

Extended Thinking being specifically valuable for decisions with real consequences is the right framing. Strategy, technical architecture, pricing scenarios where the cost of a shallow answer is high enough to justify the extra time. Using it on everything produces slower results without proportionally better outputs.

Projects with uploaded knowledge applying automatically across every conversation in that project is the workflow architecture most users set up once and then stop noticing because it just works. What knowledge have you put into Projects and what are you still reprompting from scratch each session?