Long context is turning software into a place, not a tool

Capacity is no longer the whole problem

Anthropic made a one-million-token context window generally available for Claude Opus 4.6 and Sonnet 4.6 at standard platform pricing. That is enough capacity to place large codebases, document collections, or extended project histories in a single working context.

Earlier, the jump from 9,000 to 100,000 tokens was framed as the ability to process roughly 75,000 words. Each order-of-magnitude increase changes what is technically possible, but also raises a harder product question: what should the system remember, retrieve, summarize, or forget?

People need continuity, not a bigger bucket

A human collaborator does more than retain every sentence. They carry goals forward, recognize decisions, separate durable facts from temporary details, and ask when an old assumption no longer applies. Product memory has to do the same work visibly enough to earn trust.

Long context can reduce retrieval boundaries, but it can also increase latency, cost, and the chance that irrelevant history distracts the model. Good interfaces give people control over the working set instead of presenting memory as magic.

The interface becomes a place

When a system maintains projects rather than isolated prompts, navigation, provenance, and state become central. The next leap is not another number in a model card. It is helping people understand what the assistant knows, why it surfaced something, and how to correct it.