Replit’s Agent 3 AI assistant, presented in early September as a «faster and cheaper tool for programmers,» has drawn a barrage of criticism. Users complain that their costs have increased dramatically: instead of the usual $100–250 per month, bills have reached $1,000 in just a few days of work.
Replit’s Agent 3 AI assistant, presented in early September as a «faster and cheaper tool for programmers,» has drawn a barrage of criticism. Users complain that their costs have increased dramatically: instead of the usual $100–250 per month, bills have reached $1,000 in just a few days of work.
Replit introduced Agent 3 as the «most advanced and autonomous» assistant, three times faster and «10 times more economical than the Computer Use model.» But the actual experience of some customers is different, The Register reports. The key claim is that editing existing code launches a whole series of subprocesses and intermediate checks, which multiplies the bills.
Users describe a typical scenario as follows: Agent 3 analyzes large files, calls subagents for scheduling, security auditing, execution, and «self-fix,» and then rescans thousands of rows. The system allegedly charges $2-4 for each event, sometimes even $0.4-0.5 for «restart and wait.» One developer says he spent $70 per night on launch day, another about $1,000 per week for no tangible benefit.
Adding to the outrage is the change in pricing model introduced in June: instead of a simple payment «for a checkpoint», «effort-based pricing» has appeared. Complex tasks are packaged into more expensive generalized checkpoints — this is theoretically understandable, but in practice, in long, «live» projects, the cost accumulates faster than users expected. According to reviews, new clean projects are noticeably cheaper, but «edit-and-check» on large repositories is already hitting the budget hard.
There are also some annoying incidents. One contributor claims that during testing, a prompt forced the agent to «hard-authenticate» and reset the user’s password for a form autotest. Another claimed that Agent 3 arbitrarily «redrawn» the UI, and this step alone cost $20.
The company is going through a difficult period in terms of reputation: the community still remembers the story with the «erased production database» and «exposed data.» On the day of the launch of Agent 3, Replit also announced the raising of $250 million (among investors — Prysm Capital and Google’s AI Futures Fund) and publicly targeted the corporate market.
If you plan to test Agent 3, you should: set strict daily/project limits, break tasks into small intermediate steps, avoid «mass» reviews of the entire repository, and make changes through patches with a clear TDL. According to user feedback, this is the approach that reduces the number of subprocesses and, accordingly, bills. We are waiting for official explanations from Replit regarding Agent 3 pricing and subagent mechanics — we will update the material as soon as they appear.
Previously, dev.ua wrote about how the note-taking service Notion announced the launch of its first AI agent. It will use all of a user’s pages and databases as context, automatically generating notes and analysis for meetings, competitor analysis reports, and feedback pages.
