The Myth of the Silver Bullet
Educators often seek a single model for all institutional needs. In practice, commercial LLMs carry starkly different trade-offs between context windows, token pricing, and reasoning stability — a model that excels at one task can quietly underperform at another.
Soft-Caps vs. Hard-Caps
Two very different failure modes hide behind the phrase "the model stopped working":
- Hard-caps — explicit constraints where the API or web interface cuts off tokens outright, forcing a session restart.
- Soft-caps — the far more insidious, silent degradation of response quality, nuance, and instruction adherence as the context window fills.
The Token Economy
Every input and output consumes a finite context budget. Prompts are heavily subsidised in commercial tiers, but cognitive drift creeps in when older context tokens are silently discarded — the model keeps answering confidently even as it forgets what it was told three screens ago.
The Multi-Model Practice
This is why institutions need a limits-testing workflow rather than a single vendor relationship — benchmarking the leading frontier models side by side to select the right engine for each pedagogical task, rather than discovering the limits live in front of a class.