Skip to content
arter.dev

AI Code Ballin On A Budget

My ongoing experiment with an alternative workflow for maximum value and minimal cost

Chris Arter Published 4 min

I’m not freelancing right now, so I cannot justify my $200 / month Claude Code Pro Max subscription. I have been experimenting with trying to squeeze the maximum quality code output of the cheapest harness.

Note: I have decided not to use an always-free model router as these can be fickle, slow and worth what I paid. I will opt for paid inference only as I do have some standards (though not many).

Harnesses

I have evaluated three harnesses and I can share where I landed.

OpenCode

OpenCode is the most polished, fully featured alternative coding harness I’ve used. It is near parity with Claude Code in terms of developer experience and features. That said, my personal experience was that it gobbled up tokens like crazy. I concede I probably did not optimize it whatsoever. I plugged in an OpenRouter key, selected DeepSeek V4 Pro and let her rip. It was capable, but usage was just too high.

Pros:

  • Fully featured
  • Huge ecosystem
  • Great developer experience

Cons:

  • With OpenRouter it was token usage heavy.

Pi

The Pi Harness was honestly a joy to work with. It takes a minimal bring-your-own..everything approach. I was able to bolt on an agent orchestration plugin , an mcp plugin and also wire in my own static gate checks. It was extremely easy to extend and customize. I ended up using DeepSeek’s API directly with Pi code and my usage costs were dramatically lower, but probably due to other tools I’ll detail below.

Pros:

  • Great developer experience
  • Infinitely extendable, tailored to you exactly
  • No magic, simple structure

Cons:

  • Ecosystem is a bit smaller
  • You have to bolt on a lot of things (could be a Pro though?)

Reasonix

Reasonix was by far the cheapest out of the box with zero configuration. Reasonix is DeepSeek’s native coding harness tool with impressive caching performance optimized for their models baked right in. It took the least amount of work to go from install to coding efficiently.

Pros:

  • Cheap
  • Man, it’s like really cheap.

Cons:

  • Skill / tool calling is pretty poor
  • The same model and same skills in Pi fired far more reliably
  • Tiny ecosystem.

I wrote and published my own port of Superpowers but for reasonix here: https://github.com/christopherarter/superpowers-reasonix

I included multiple eval benches, tests, the works. The skills in that repository still do not invoke reliably, and I’m thinking it may have something to do with the harness itself.

Providers

So far I have used two providers.

OpenRouter

I honestly love OpenRouter for its ease of use and simplicity. I plug it in everywhere. Claw style bots, pipelines, etc. It’s broadly supported and very easy to use. That said, if you’re using DeepSeek models (which I mostly do), then I don’t think the caching performance is as good as DeepSeek’s API.

DeepSeek API

As of late, this has been the value king. I’m seeing my lowest prices (and fastest speeds?) by using the platform API directly.

Tools

I focused on these areas when configuring my local tooling

  • Context Management
  • Documentation
  • Verification & gating

Context Management

Here are the tools that are working well for me:

RTK

rtk gain --weekly

W Weekly Breakdown (3 weeklys)
══════════════════════════════════════════════════════════════════════════════════
Week                      Cmds      Input     Output      Saved   Save%     Time
──────────────────────────────────────────────────────────────────────────────────
06-08 → 06-14              726     732.8K     253.4K     479.9K   65.5%     2.6s
06-15 → 06-21             1827     837.3K     325.4K     513.9K   61.4%    957ms
06-22 → 06-28              808     953.3K     349.9K     604.3K   63.4%     1.0s
──────────────────────────────────────────────────────────────────────────────────
TOTAL                     3361       2.5M     928.7K       1.6M   63.3%     1.3s

RTK is a cli output compression tool that sits between the agent and its bash calls for common calls like git diffs and such. Grain of salt: I have to trust this data on face value because I simply don’t have the bandwidth to deep dive, but I hope it is correct!

context-mode has been useful for managing the output of noisy MCPs for token savings.

Documentation

No surprise, but I rely on context7 cli not MCP (I am generally moving towards CLI where possible).

Honorable mention, while not directly related to documentation, I run a self-hosted instance of Crawl4AI which replaces Firecrawl for my agents to grab docs, view websites, or anything else they need.