Skipping a dedicated classifier to brute-force 1,200 square kilometers across 48 vision agents is an aggressive move. Feeding images one by one through general multimodal models with a prompt guideline flagged 8 to 12-meter-high pyramid structures in under 72 hours. I think this pattern of stacking high concurrency on general models for coarse screening holds up completely.
As long as per-call token pricing stays cheap, vertical projects labeling data for custom vision classifiers will struggle to secure budgets. That money will flow straight into inference compute for general multimodal models instead. Anyone whose playbook relies on selling specialized small models is going to feel real pain.