Charging by points for the warp agent is better than charging per use. If properly tuned, it will outshine all AI IDEs that charge per use. End with key cost-saving tips.

Tested: 1000 lines of ASP backend files, let GPT-5 refactor and split, which is a heavy task with high demands on the model's intelligence and speed. GPT-5 is extremely fast, so fast that it surpasses the experience of Claude Sonnet4, while also exhibiting very high intelligence; it feels like a professional backend has joined.

Points consumed: 16/10000
Subscription plan: $50, similar software, such as augmentcode, offers 600 credits.

Key points to save money:

The points system of warp is dynamically billed based on your token consumption, when your demand composition is

High Frequency: Refinement (fine-tuning button styles, function writing), only consumes about 5 points, actual measurement is 4 points = 2500 times.
Low frequency: refactor code (split large files for refactoring), only use 20 points = 600 times
Ultra Low Frequency: Read the complete project document rules, provide professional advice, 50 points = 200 times

It means that a normal programmer, who buys 10,000 points for $50, will use them far more than 600 times compared to competitors, and even cursor isn't worthy of tying their shoes.

Beware of token assassins:
The GPT series is very economical with its output. If you follow a conventional conversation, it will read some files in a few steps, then start talking about a lot of its plans below, asking you if you want to continue, and fulfill one of your requests. In this way, the conversation can consume more than fifty times of the quota.

Saving Password (Insert Mandatory Prompt): You complete all the work at once and do not need to ask me if you want to continue.

16 = 200 times => 625 times