After using OpenClaw for a few days, I was on the verge of bankruptcy.

The article shares personal experiences with AI tools, focusing on cost and usability issues.
High-end models like Claude Opus are intelligent but expensive, while cheaper models like DeepSeek are free but limited in performance.
Subscription-based services conflict with pay-per-use mentality, and switching models disrupts the experience.
OpenClaw is favored for its human-like interactions.
Improper context management wastes resources, especially when using costly APIs.
Various features depend on APIs, increasing additional expenses.
Security concerns are manageable, but caution is advised for sensitive information.
In summary, using AI is seen as a micro-investment, akin to gambling, with hopes for high-quality output.

Direct feeling: On the verge of bankruptcy.

1. Are there really different levels of rank among employees?

At first, I used Claude Sonnet 4.5, and it really made spending money (tokens) like water. One sentence cost one dollar, and I used up all my tokens in no time.

It's hard to go from luxury to frugality. Later, when I switched to a cheaper model, I felt that the intelligence level was too low to accept (for example, Claude 3.5).

You get what you pay for; if you pay ten times the price and get only a fraction of the quality, it's hard to choose.

2. Monthly subscription mindset vs. per-item mindset

Being good at managing money is certainly a good quality among older Chinese people. However, the downside is that when I want to discuss certain issues, I always prefer to look for Gemini and GPT with monthly memberships rather than OpenClaw, since they operate on a per-item basis.

However, this constant switching is really ruining the experience, to be honest. I feel like an old person from the past who has to turn off all the lights and air conditioning before going out for dinner.

3. It feels like a human being, and like AI.

I think one thing people like about OpenClaw is that it speaks more like a human. GPT Gemini, on the other hand, if you haven't programmed it, deliberately behaves like an AI. I suspect that OpenClaw incorporates a human-like persona into its settings, which has won over many people who don't like programming AI.

4. Domestic products are always plentiful and satisfying.

One major solution to token anxiety (people are really doomed these days; mobile phones cause charging anxiety, electric vehicles cause range anxiety, and now even AI has token anxiety) is to switch to DeepSeek. It's free and you can use it without worry.

Domestically produced goods are always plentiful and satisfying, but in terms of intelligence, they certainly lag behind Claude Opus 4.5 and 4.6. Moreover, they are somewhat limited in terms of certain values.

The price difference is 20 times... Can you believe it?

5. Hybrid electric vehicles are urgently needed.

I urgently need a heavily modified version of OpenClaw, which is a dual-model system, similar to new energy vehicles and hybrid electric vehicles.

- For casual chatting, checking literature, and completing daily tasks, use a cheap and plentiful model;

- But if I need to write code, automatically switch to the most expensive Claude Opus 4.5/4.6 and output the actual usable code, not toy code.

Use a chicken knife to kill a chicken, and an ox knife to kill a cow.

6. Use with caution if you have a strong aversion to using certain contexts.

I have a strong aversion to contextual cleanliness. Ever since I learned that every additional piece of context for an AI would multiply the computing power required, I have particularly disliked completing all tasks in a single dialog box. If the tasks are unrelated, I will not hesitate to open a new dialog box.

However, OpenClaw doesn't seem to give me a chance to reopen the dialog box. It's really frustrating to watch my dear assistant thinking about 100K/200K of contexts every time.

Because this context includes all sorts of random chatter, failed tasks, and questions about OpenClaw operation commands—all this random stuff is mixed up with my Vibe Coding requirements... especially when using the expensive Opus 4.5 API, it's heartbreaking.

7. Everything requires an API

Everything requires an API. For example, even search requires an API. I bought Brave Search AI, but the free version wasn't enough; I had to pay for it. Even getting Bitcoin prices requires the Coingecko API. In short, there's a lot of large-scale purchasing and consumption.

8. Security issues

The security issues aren't as serious as I imagined when I was a cloud user, especially since I was a bit overly worried about injection attacks. It also has quite a few security guidelines of its own. However, I still won't put important assets and passwords on it.

at last

The initial claim of being on the verge of bankruptcy was a joke, and it's not that expensive. But honestly, each dialogue generation is actually a "micro-investment," one that you can invest with your money and time.

There's an old saying about AI that goes something like this: every time you burn tokens and time to do Vibe Coding, you're playing a slot machine. You're betting that it will produce a perfect piece of code, betting that it's not a pile of junk that can never be fixed.

Wishing everyone good luck in gambling.