The model launched on June 9 arrived with a one-million-token context window and always-on adaptive thinking, both announced as differentiators on long and complex tasks.
A million-token window means, in practice, putting an entire codebase, or hundreds of pages of documents, inside a single conversation. That changes application design: part of the work of selecting the relevant excerpt before asking stops being necessary.
The gain comes with a direct cost, and the arithmetic is simple. A large context is billed per token processed, and sending a million tokens with every question is expensive enough to rule out repetitive use. The window serves the task that justifies the spend, not the routine.
Always-on reasoning has a similar effect: it improves results on hard problems and makes simple answers more expensive, reinforcing the model-routing logic that became standard in the period.
The ending is worth recording: three days after launch, the model went offline by government order, and anyone who had designed an application around that window lost the central piece.