Esc
Type to search posts, tags, and more...
 Skip to content
3 min read

Using tokens economically for hobby projects

API access to frontier models stopped making sense after my AWS credits expired. For hobby projects, subscriptions and provider-neutral agent harnesses are a better deal.

--ai--tooling
Contents

For a long time, I did not have to think much about token costs. I had a pile of AWS credits, and Bedrock gave me access to Claude Opus 4.6. It worked well. The credits made that level of model usage feel almost unlimited.

Then the credits expired.

Paying the API price for large models is hard to justify for hobby projects. Once every agent run has a visible cost, experimentation feels different. A long session, a few subagents, or an unproductive direction can turn into real money quickly.

Subscriptions changed the calculation

I started looking for a flat-price replacement.

Claude Pro was the first option, but its limits felt too restrictive for the way I work. I also tried ChatGPT Plus, which has been a much better fit. I use it with Codex and the Pi agent, and I have not run out of the provided usage window on my hobby projects.

That costs $20 per month. It also includes the ChatGPT mobile and desktop apps. I do not know whether this pricing is profitable for OpenAI or how long it will stay this generous, but the value is good today.

GPT-5.6 Sol with medium reasoning has been enough for almost anything technical when paired with the right workflow. The main thing I still miss is access to an Opus-class model.

I still want access to Opus

I want to keep testing Opus, so I will likely add Claude Pro again and use it through Claude Code. One workflow I want to try is using Pi as the main agent and invoking Claude Code as a subagent when I specifically want Opus.

That gives each subscription a role:

  • ChatGPT Plus and GPT-5.6 Sol for most technical work.
  • Claude Pro and Opus for comparison and tasks where I want a second frontier model.

This is not about finding one permanent winner. It is about getting useful access to different models without paying frontier API rates for every experiment.

Kimi adds another option

Kimi K3 is another interesting model. It is advanced, relatively cheap, and a Kimi subscription could cover more AI usage at a flat price.

There are now several ways to assemble an economical setup:

  • Use subscriptions for regular interactive and agent work.
  • Use OpenRouter when a specific model is only available through an API.
  • Spend cloud credits where they are still available. I still have Google credits, so I use those for Gemini.

The exact mix will change. Model capabilities, prices, and subscription limits are moving too quickly to assume that today’s setup will still be the right one in six months.

Do not lock the workflow to one provider

This is why I do not want to invest too deeply in a single agent harness or provider-specific toolchain right now. The useful models keep changing. So do the cheapest ways to access them.

For hobby projects and small startups, the economical approach is to use the value already bundled into subscriptions, keep API access as an escape hatch, and choose tools that let you switch providers quickly.

The model is replaceable. The workflow should be too.

! Was this useful?