AI engineering is not picking a model; it is designing the system around the model. What goes into the context and what stays out, how a failed call is retried, where token cost can be cut, and how you find out later what happened all belong here, and those decisions stay when you switch models.
The reports work from measured data. All 18 models tested got worse as inputs grew longer, well before reaching their limits. Across 12 conditions, caching only the system prompt was most often the best: on cost for 3 of 4 models, and on time to first token for 3 of 4. A timeout is not a confirmed failure but an unknown outcome, which shapes retry design. And every OpenTelemetry GenAI convention is still at the Development stability level.