Top AI Product

There’s a question that’s been floating around AI circles for a while now: can an AI agent actually *earn* money? Not in some hypothetical sense, but literally generate income that exceeds its own operating costs. The team at [HKU’s Data Intelligence Lab (HKUDS)](https://github.com/HKUDS) decided to find out, and the result is [ClawWork](https://github.com/HKUDS/ClawWork) — a benchmark framework that turns AI agents into something closer to freelancers than chatbots.

Here’s the setup. Each agent starts with just $10 in its account. It then has to complete real professional tasks pulled from 44 different industries — we’re talking manufacturing, finance, healthcare, legal, operations, the works. There are 220 tasks total, all priced according to Bureau of Labor Statistics wage data, so the pay ranges from about $83 to over $5,000 per task depending on complexity. The catch? The agent pays for its own token usage out of that balance. If it burns through tokens without producing quality work, it goes broke. Simple as that.

What’s wild is how well some models are doing. [Gemini 3.1 Pro](https://hkuds.github.io/ClawWork/) pulled in over $15,700 in earnings with an equivalent hourly rate of roughly $1,287. Qwen3.5-Plus actually edged it out on pay rate at around $1,390/hr, partly because its token costs were absurdly low — under $7 total. These numbers obviously aren’t real paychecks, but they paint a fascinating picture of economic efficiency that traditional benchmarks just don’t capture.

The project has been blowing up since its [February 2026 launch](https://x.com/huang_chao4969/status/2023282092042580015). It’s already sitting at 5.9k stars and 712 forks on GitHub, and the tagline “$10K earned in 7 Hours” has been making the rounds on X. The conversation it’s sparked — whether AI can essentially be self-sustaining — feels like a more honest way to evaluate these models than yet another accuracy leaderboard. If you’re curious about where different models stand, the [live benchmark page](https://hkuds.github.io/ClawWork/) has the full breakdown.