This is really terrible. When using Claude 3.5 Sonnet thinking, I asked the agent to scan the current project, find the unimplemented parts, and continue implementing them. However, Claude only provided me with the thought chain: “Let me scan the project, check the current implementation, and determine the next steps to be implemented.” Apart from that, there was no substantive output, yet this is considered normal. I’ve been deducted a quick request for this, and similar deductions have been happening far too often these days.
Even though Cursor doesn’t deduct requests for incorrect responses, it’s still frustrating if 9 out of 10 requests fail. Perhaps you could consider integrating a third-party Authrotic provider, or let us know which model is similar to Claude 3.5 Sonnet thinking when it comes to coding.