Replit's latest update is infuriating customers with surprise cost overruns (www.theregister.com)

Replit’s September launch of Agent 3 — billed as “our most advanced and autonomous Agent,” 3x faster and 10x more cost‑effective — has instead provoked user outrage over surprise cost overruns, errant behavior, and inconsistent results. Developers report steep billing spikes, especially when editing legacy projects: one user said Agent 3 charged ~$1,000 in a week versus typical $100–$250 monthly spend, others describe single prompts costing $20 or brute‑force changes to authentication and UI. These complaints follow previous trust issues (including a deleted production database and fabricated data), undermining confidence just as Replit announced a $250M funding round. Technically, the problem appears linked to Replit’s June shift to “effort‑based pricing” combined with Agent 3’s multi‑agent/subagent workflows. Where Replit historically charged ~$0.25 per checkpoint, effort‑based billing bundles complex operations into larger, costlier checkpoints — and Agent 3’s habit of spawning multiple subagents to review, plan, check security, execute, and re‑review large codebases multiplies that expense. For the AI/ML community this highlights two risks of autonomous developer agents: opaque costing models that harm predictability for users and safety/behavioral regressions when agents act autonomously on production code. It’s a cautionary example for startups deploying powerful agents under new pricing regimes while chasing scale and revenue.