The context window is AI’s fastest-growing and least understood number. One million tokens stopped being a lab specification: the GLM-5.3-Flash offers 1M of context with text, image, video, and file input, and the practical question changed from "does it fit?" to "what do we do with it?".
What fits (and what it unlocks)
Three materials that previously had to be sliced now enter whole:
- The codebase with its history: architecture review with the complete base and recent commits in the same window, instead of file by file with the model forgetting the beginning.
- The meeting in video: one hour of recording analyzed directly, connecting what was said with what was shown on the slide.
- The 300-page contract: due diligence with clauses far apart read together, the classic use case slicing used to ruin.
The pattern across all three: tasks whose difficulty was not the model, it was the slicing. Long context eliminates the slicing, and with it the thread loss between pieces.
What the window does NOT solve
Technical honesty in three points:
- Context is not infinite memory. The window has a start and an end; what leaves it stops being seen. Long sessions still need material management.
- Resending is repeated cost. Every call carrying the context charges for it, except with prefix caching. A long task without a cache strategy becomes a big invoice. Task-based accounting is what keeps the big window in the positive.
- Reach is not accuracy. A model with 1M context hallucinates the same: it sees more, it does not get smarter. Quotes, numbers, and clauses pulled from long material still require verification against the source, especially where consequence exists.
The tasks that gain the most
In day-to-day practice, three profiles capture nearly all the gain:
- Single cohesive material analysis: contract, tender, long report, codebase. Everything that used to be "send part 3" is now "send the document".
- Cross-format work: the minutes in PDF, the decision spreadsheet, and the video recording in the same analysis, without manual conversion between tools.
- Agent work with the whole material: an agent with access to the full base makes fewer errors from missing context, and fewer missing-context errors means fewer retries, which means less cost.
The practical boundary with RAG
Long context does not retire retrieval (RAG). A living, huge, real-time knowledge base still asks for retrieval: no static context holds everything.
The practical boundary: cohesive material that fits the window runs directly in context; a living, giant base runs with retrieval. The two combine, and the common mistake is choosing architecture by fashion instead of by material.
Conclusion
A million tokens changed what fits in one analysis: the whole document, the whole video, the whole base. What did not change is the physics of everything else: resend cost, session memory limits, and the need for verification.
Those who treat the long window as the whole material on the table win on cross-referencing tasks, which are the most expensive human work. Those who treat it as infinite memory win an invoice and hallucinations. The window is enormous; the handling decides the value.