The Subsidy Is Over

OpenClaw and tools like it run on a logic that made complete sense at the time. You just make it do things continuously, on demand, for anything. The reason it worked like magic was not due to architectural genius but the fact that Anthropic and OpenAI were effectively subsidizing the compute. Flat-rate subscriptions decoupled usage from cost. Users ran agents hard because they had no reason not to.

Running frontier models at full capacity for an open-ended workload is expensive. The economics only worked while the providers were choosing growth over margin.

Well, that period is ending. Major frontier model providers are tightening, through usage caps, rate limits, or direct price increases. The era of treating a powerful model as a free resource is closing. OpenClaw, in its original form, was a product of a specific market condition. That condition is changing.

The agents that survive this shift are the ones built with efficiency as a constraint from the beginning. That means tightening the loop by choosing the right-sized models for the task at hand, instructions optimized to minimize unnecessary inference, tools granted only where they add real value rather than out of convenience. For parts of a workflow that do not require frontier-level reasoning, open-source models cost a fraction and perform well. Mixing model tiers based on task complexity is the obvious move, and most agent builders have been slow to do it.

Agents built on the assumption of cheap, unlimited frontier inference will get repriced into irrelevance. Agents built with efficiency as a core property survive and become more competitive as the gap widens.

A practical note: ChatBotKit is built around this model. The platform lets you configure the right model per use case, compose optimized instruction sets, and attach only the tools each agent actually needs - so you are not running a frontier model at full throttle for tasks that do not require it.