Last week we discussed OpenClaw and the world of AI agents. If you missed it, you can catch up on the blog here. The short version: AI that doesn't just answer questions, it actually does things on your behalf. Schedules meetings. Clears your inbox. Runs while you sleep.
Sounds great. Then the invoice shows up.
My first full day running my own agent cost me $20 in Anthropic API credits. One day. That's when I went from excited early adopter to very attentive accountant. I started mapping out every task my agent handled and asking a question I'd never really had to ask before: does this specific job actually need the smartest, most expensive model available?
Usually, the answer is no.
Most people never think about token costs when using ChatGPT or Claude through their normal apps. You pay a flat monthly fee and move on. The moment you start building agents and calling the API directly, that changes. You're paying per token. A token is roughly three-quarters of a word. Every word your agent reads, every word it writes back, every instruction you've given it, all of it counts. A busy agent doing real work burns through tokens fast.
This is why Qwen, DeepSeek, and GLM are showing up everywhere in agent-building circles. These are Chinese-developed models, and they're genuinely good. Not "good for the price" good. Just good. DeepSeek caused real waves last year when it matched GPT-4 level performance at a fraction of the cost, which led a lot of people to wonder whether the big US labs had been massively over-engineering their way to results. Lower infrastructure costs, more efficient training approaches, and strategic pricing to gain market share all play a role. Whatever the mix, the capability-to-cost ratio is hard to ignore.
A few people have reached out lately feeling like they're falling behind. Their feeds are full of OpenClaw and AI agents and they want to know what they're missing. My candid take: I've spent considerably more time configuring OpenClaw than it has saved me so far. It's still a hobbyist tool. Rewarding if you enjoy tinkering, but not something most people should feel pressure to adopt right now.
That said, this space is moving fast, and we want to make it easier to follow. We just launched the LA-AI Agent Directory at la-ai.io/directory The directory uses an AI agent to do daily scans for new or updated agents and their accompanying support tools. Additionally there are a series of documents to help people get familiar with some of the key concepts and how-tos.
New agents are coming online every day, along with the supporting tools you need to actually run them, and keeping track of it all is its own job. The directory is being built and maintained by an AI agent, which feels appropriate. It's a good place to stay current without having to live on GitHub or Reddit (that’s my job, after all).
One final note. If you really want to get into it, check out the #claw-mayhem channel on the LA-AI Discord server https://discord.gg/9HKU4GAGDC