│ > Try this again using mpfr, which is already installed. Call this │
│ pi_value_mpfr. │
Gemini got that right the first time, which wasn’t a huge surprise since MPFR has a built-in constant function for pi. Then Gemini volunteered that the MPFR implementation was a bit slower than the Boost implementation. Well, I knew that MPFR was using a suboptimal but reliable algorithm.
Me:
│ > try again using the best possible algorithm. call the program │
│ pi_value<algorithm_name> │
Gemini correctly chose the Chudnovsky algorithm with MPFR, but then screwed up the implementation a few times before getting it right. When it finally got a correct result, the real runtime was about half of that of the built-in constant function for pi. I then had Gemini check the three result values against each other and the 39-digit reference value given in the Encyclopedia Britannica. All good.
On to something non-trivial: I changed to the directory where I had a current copy of the gemini-cli repository and first asked it to explain the repo, and then asked it to list open issues, which it did using gh once I gave it permission. Then I asked it to pick an open issue marked as a good first issue, and then to suggest an implementation. It came up with a plan, ran into trouble with the implementation, came up with a revised plan, had trouble with that, and finally went into a loop trying to balance TypeScript parentheses. I canceled the task at that point.
Good but not great
My conclusion here is that Gemini CLI with the gemini-2.5-pro model is pretty good but not great. On a scale of Level 0 to Level 5 where Level 0 is “No automation or agency, but good tooling for shortcuts and repetitive use” and Level 5 is “Full automation: Software can problem-solve and implement without supervision,” the Gemini CLI and Pro model are probably at around a 3.5, somewhere between Level 3’s “Conditional automation: Software can write itself under pre-defined rules and recipes” and Level 4’s “High automation: Software can write itself with little to no human supervision.”
To put that in perspective, the only better systems that I’ve tested are Claude Code and Solver, which can both sometimes behave at or near Level 4. On the other hand, Gemini CLI is free and the gemini-2.5-pro model has a million-token context window and advanced coding capabilities, which together can get you pretty far without paying for a subscription or for usage.
Cost
Free with 60 model requests per minute and 1,000 requests per day; usage-based billing and subscriptions available.
Platform
Command line/terminal; Node.js v18+; Windows, macOS, Linux.
Bottom line
Gemini CLI is free and the gemini-2.5-pro model has a million-token context window and advanced coding capabilities, which together can get you pretty far without paying for a subscription or for usage. The only better systems that I’ve tested are Claude Code and Solver.
Pros
- Free for most uses
- Good results most of the time
- Somewhere between conditional and high automation
- Active open-source project
Cons
- Still buggy since it has been available for less than a month
- Uses a hosted model, which means that your code could potentially leak
- When it fails it can produce garbage code but claim that it works
