│ > Do this once more utilizing mpfr, which is already put in. Name this │
│ pi_value_mpfr. │
Gemini acquired that proper the primary time, which wasn’t an enormous shock since MPFR has a built-in fixed perform for pi. Then Gemini volunteered that the MPFR implementation was a bit slower than the Enhance implementation. Properly, I knew that MPFR was utilizing a suboptimal however dependable algorithm.
Me:
│ > strive once more utilizing the absolute best algorithm. name this system │
│ pi_value│
Gemini accurately selected the Chudnovsky algorithm with MPFR, however then screwed up the implementation a couple of instances earlier than getting it proper. When it lastly acquired an accurate outcome, the true runtime was about half of that of the built-in fixed perform for pi. I then had Gemini verify the three outcome values towards one another and the 39-digit reference worth given within the Encyclopedia Britannica. All good.
On to one thing non-trivial: I modified to the listing the place I had a present copy of the gemini-cli repository and first requested it to elucidate the repo, after which requested it to record open points, which it did utilizing gh as soon as I gave it permission. Then I requested it to select an open subject marked as a good first subject, after which to recommend an implementation. It got here up with a plan, bumped into bother with the implementation, got here up with a revised plan, had bother with that, and at last went right into a loop attempting to stability TypeScript parentheses. I canceled the duty at that time.
Good however not nice
My conclusion right here is that Gemini CLI with the gemini-2.5-pro mannequin is fairly good however not nice. On a scale of Stage 0 to Stage 5 the place Stage 0 is “No automation or company, however good tooling for shortcuts and repetitive use” and Stage 5 is “Full automation: Software program can problem-solve and implement with out supervision,” the Gemini CLI and Professional mannequin are most likely at round a 3.5, someplace between Stage 3’s “Conditional automation: Software program can write itself below pre-defined guidelines and recipes” and Stage 4’s “Excessive automation: Software program can write itself with little to no human supervision.”
To place that in perspective, the one higher methods that I’ve examined are Claude Code and Solver, which may each generally behave at or close to Stage 4. Alternatively, Gemini CLI is free and the gemini-2.5-pro mannequin has a million-token context window and superior coding capabilities, which collectively can get you fairly far with out paying for a subscription or for utilization.
Price
Free with 60 mannequin requests per minute and 1,000 requests per day; usage-based billing and subscriptions out there.
Platform
Command line/terminal; Node.js v18+; Home windows, macOS, Linux.
Backside line
Gemini CLI is free and the gemini-2.5-pro mannequin has a million-token context window and superior coding capabilities, which collectively can get you fairly far with out paying for a subscription or for utilization. The one higher methods that I’ve examined are Claude Code and Solver.
Professionals
- Free for many makes use of
- Good outcomes more often than not
- Someplace between conditional and excessive automation
- Lively open-source undertaking
Cons
- Nonetheless buggy because it has been out there for lower than a month
- Makes use of a hosted mannequin, which implies that your code might doubtlessly leak
- When it fails it will possibly produce rubbish code however declare that it really works
