Agents are expensive, like awfully expensive. For example, Claude can charge up to $25 per million tokens, almost the same, but a little bit cheaper, for Codex from OpenAI. So in rough calculations, depending on usage, you can spend up to $150-170 a month (if you check Reddit, people proudly burn $2k-4k a month, this is insanely expensive). Cloud Code is powerful, possibly the best agent out there. The problem is, it drains my $20 PRO usage in a few runs. With agents, you might accidentally burn all tons of tokens if you're not careful enough, and this is a major problem keeping me away from switching to the pay-as-you-go plan. Skills also affect token usage and bump the usage up to 50% more. Not even talking about the MCPs.
For the last month I was using agents and workflows to understand the most cost-efficient way to fulfill my needs for my side hustles. Currently, I use two Agents (probably will cancel one in the future) and one LLM. My setups is the following:
With Opencode and Junie, I use $10 subscriptions (roughly converts to 10 credits on each of them). Believe me, this is a huge amount if you use it correctly. For example, I still write code manually occasionally, and mostly use agents for code reviews, code updates, generating tests, improving coverage, and performing partial refactorings in a small scope of the code.
Importantly, NEVER ASK OR PROVIDE ABSTRACT TASKS OR QUESTIONS. Since I'm the owner of the code, I know what and where it is in my project. Instead of injecting a prompt to the agent to go and find X and perform Y, I explicitly guide them:
"Add additional unittests for X. You can find the already existing Y. Package is missing unittests for [L:100-120] and [L:300-310] when calling thisImportantMethod"
This saves cycles, and the agent avoids calling additional tools that cost tokens. One more important thing: ALWAYS /init your project with AGENTS.md. This will save you tokens on every new session.
What about Claude Sonnet 4.6? I needed a second brain for brainstorming. Sonnet 4.6 is enough to discuss an idea with additional research notes from an LLM (it also provides gorgeous diagrams). It has decent power and is not that token-hungry like Opus.
So, whenever I do something, like POCs or improvements to my existing projects, my approach is:
As a reference, with DeepSeek 4 Flash I spent $6.11 worth of tokens. On Junie I spent around 72% of my credits, 7.13/10.00 credits. The load was about 20-25 tasks/day.