iTranslated by AI
My Impression of OpenHands (OpenDevin): "Devin is Actually Cheaper"
I really liked using Devin for work, so I wanted to use it for personal projects too.
However, spending $500/month for personal use is frankly impossible, so I turned my attention to the OSS OpenHands (formerly OpenDevin).
But after using it with Sonnet 3.7 (3.5), I actually came to think, "Isn't the original Devin actually cheap?" So, I'll write about what I did and my impressions.
About OpenHands
OpenHands is an open-source AI agent that assists with engineering tasks. It is known as an open-source implementation of Devin (formerly called OpenDevin) and supports various development work.
It is said to be capable of things like launching new projects, adding features to existing codebases, and refactoring. For example, it seems capable of handling a wide range of tasks such as creating new React components, adding GitHub actions, and fixing bugs.
However, as many engineers have already pointed out, current AI agent technology (not just OpenHands) cannot fully automate complex development tasks. Human guidance is still necessary, and one must be mindful of things like "breaking tasks into smaller units" and "providing concrete instructions."
Devin is expensive... so OpenHands!
Devin is priced from $500 (¥75,000)/month, which is frankly difficult for personal use.
So, I decided to try using OpenHands to see if I could use an AI agent at a lower cost.
OpenHands is basically assumed to run on a local machine. However, I had a need to "AI code from my smartphone even when I'm out," so I decided to deploy it on Google Cloud with the following configuration.
Google Cloud Architecture Diagram
As a result, I was able to use it from both PCs and smartphones via the internet, just as I intended.
Using OpenHands deployed in the cloud
Now I can AI code from my smartphone even when I'm out
About $30 (¥4500) melted away in about 1 hour 😂
I started with Anthropic's API key, but frequently encountered rate limit issues. So, I paid $40 to upgrade to Tier 2 (for Tiers and rate limits, see Anthropic's documentation).
However, I continued to hit rate limits, so I switched to OpenRouter's API key.
But Claude Sonnet 3.7 (3.5) is quite expensive 😇 It's well known that using the Claude API extensively in contexts like Cline can incur significant costs.
In the case of OpenHands, when I tried to have it perform coding tasks conversationally for personal development, about $30 melted away in about an hour. Below is a capture from https://openrouter.ai/activity:
About $30 was consumed quickly
Approximately $0.4 (¥60) per request 😇
And in addition to the LLM usage fees, there will also be fixed costs for Google Cloud's Compute Engine and Cloud Load Balancing.
Devin is $8 (¥1200) per hour
On the other hand, regarding Devin's cost. Although the $500 monthly fee tends to catch the eye, let's consider it on an hourly basis.
According to this, 1 ACU is equivalent to about 15 minutes of work. If we accept this, it means 4 ACUs per hour. And since 1 ACU costs $2, it comes out to $8 (¥1200) per hour. Perhaps Devin is inexpensive...?
Compared to GitHub Copilot or Cursor, which offer essentially unlimited AI usage for $10-$20 a month, this might seem quite expensive. However, as an autonomous AI agent that's not just a coding editor companion, I think this pricing is actually quite competitive.
Considering that ACU includes the instance costs for the execution environment, and there are other functional advantages such as:
- Parallel execution possible
- Also, the development environment (instance) is completely independent for each instruction
- Slack integration
- Comments on GitHub Issues and PRs
- High task accuracy
- This is subjective, but I felt Devin coded more accurately to the intent of the instructions.
- Devin seems to combine multiple AI models, letting each handle its specialty (OpenHands specifies a single model) https://x.com/ScottWu46/status/1895209597084017073
- This is subjective, but I felt Devin coded more accurately to the intent of the instructions.
- Ability to interrupt and give instructions during AI agent work
- OpenHands currently does not support this (you can't speak unless you explicitly stop the task)
Considering these points, I believe it's actually quite inexpensive.
Strengths of OpenHands
Of course, OpenHands has its strengths.
First, because it allows you to choose low-cost models like DeepSeek and Gemini 2.0 Flash, it's highly likely to be more affordable than Devin in those cases (though I still prefer Sonnet for coding...).
Also, since it's OSS, you can customize the source code as needed (if you have the technical skills and motivation) and adjust it to your development environment and requirements. As it is still under development, further improvements from the community can also be expected.
Summary
Devin might seem expensive at first glance, but considering the actual cost per hour of use and the features it provides, I found its cost-performance to be quite good. On the other hand, OpenHands can be an option if you want to reduce costs by avoiding higher-priced models, or if you prioritize customization and self-hosting.
Nevertheless, the fact that $500+ per month is quite steep for personal use remains unchanged, so I want to support OpenHands' efforts. I'd like to try code reading and contributing if I get the chance.
That's all for why I think Devin is actually inexpensive.
Discussion