OpenAI quickly faced the issue of exceeding expected quota consumption after launching GPT-5.6 Sol. Codex Head publicly acknowledged that during the pre-release phase, there was excessive focus on average and median consumption, underestimating the long-tail heavy scenarios, and immediately reset quotas, optimizing tool calls and waiting logic.

In terms of capital and product pathways, stronger autonomy (long-duration work, parallel tools, sub-agent coordination) directly translates to higher token consumption, forcing the platform to quickly address efficiency shortcomings after a leap in capability; resources are tilted towards runtime optimization, cache management, and differentiated quota design to balance cutting-edge capabilities with sustainable usage costs.

A similar path can be seen in the consumption curve corrections of previous generations of models from laboratory to large-scale deployment: capability leads, and efficiency follows is a common rhythm. The current agent workflow is still in the "parallel iteration of capability release and cost control" stage.

Essentially, this is a technological substitution: as models shift from single-round generation to multi-step autonomous tool usage, billing and limitation mechanisms must transition from static quotas to dynamic efficiency perception, with pricing power shifting to platforms that can simultaneously enhance capabilities and control consumption.

ABAB News · Cognitive Laws

Capability leaps often first appear in the form of exceeding expected consumption.
Median efficiency cannot mask the true costs of long-tail heavy users.
The stronger the autonomous workflow, the more critical runtime optimization becomes.