Cursor adds DeepSeek V4.1 Flash: 60 cents per task, and one user spent $24.60 without fixing a bug

On October 26, Cursor announced that DeepSeek V4.1 Flash was live. Open Settings > Models to use it. The company also posted a chart showing that it sits on the CursorBench cost frontier at every effort tier, at only $0.60-$0.98 per task. But the most popular item in the replies was a screenshot showing 'the login bug is still open.'

First, the official chart. The x-axis is average cost per task (USD), and the y-axis is the CursorBench 4.0 score, Cursor's own coding model benchmark. DeepSeek V4.1 Flash 1017 lands in the bottom-right corner: $0.60-$0.98 per task, covering all effort tiers from low to high. According to the company, no model scores higher in this cost range. Other spots on the chart are also worth noting: user George Shostakovych noticed that GPT-5.6 Sol sits alone in the high-cost, high-score corner, writing, 'That poor 5.6 sol is all alone'; Haiku 5.5 is also in the low-cost area, but its score is lower than DeepSeek V4.1 Flash.

Now for the screenshot, from user Harjot Rana. His Agent Session is titled 'Fix the login bug.' Key data: $0.60 per task, 41 tasks in total, $24.60 total spend, including 12 instances of 'You're right, let me revert that.' The session ended with the login bug still open and 3 new bugs discovered.

This screenshot illustrates the problem better than the official chart. The price is indeed cheap, but whether the Agent can solve your problem is another matter. Harjot's login bug turned into a marathon of 41 loops that ended with 'the bug is still there, new bugs +3.' Anyone who has used an Agent to write code should be familiar with the 'You're right, let me revert that' loop.

The community reaction falls into two camps. One camp cheers the price. User Knight said, 'At 60 cents per task, I have no excuse to keep avoiding that bug I've been avoiding since Friday'; user Blum said, 'At this cost, I can basically leave high-effort mode on all the time.' User Irien posted a video with the caption 'YOU ARE SO ON ANOTHER LEVEL.', meaning the price is so cheap that you no longer have to second-guess every request.

User trap also mentioned that previously this model could only be routed through OpenRouter, but now it is built into Cursor, which is much more convenient.

The other reaction is about practical questions. Multiple users are asking the same question: Is the model included in subscription plans, or is it billed via API? User Batatorfel asked whether Cursor has a quota for open models. User Albert asked when Composer 3 will be available. The company has not responded yet.

One signal: as an open-source model, DeepSeek V4.1 Flash has now squeezed onto Cursor's cost frontier. When the price per task drops below $1, user behavior changes: people no longer need to calculate whether every request is worth sending. But 'cheap' only eliminates cost anxiety. The Agent's own judgment and ability to backtrack are what really determine whether it is worth turning on high-effort mode.

Keep an eye out for follow-up updates.