I mostly see comparisons between models on autonomous programming benchmarks, but they are almost orthogonal to how good the models is when used more as a pair programming partner.
My sessions are basically:
- ask AI questions about codebase
- probe details until we share a solid mental model of code as–is
- explain what I want to implement and suggest some architectural choices
- have AI evaluate my suggestions and propose its own
- discuss trade-offs, finalize the plan
- send AI to implement
*Plus custom documentation system to preserve knowledge between sessions.
For this back-and-forth work, I’d take Kimi 2.7 over Claude Opus 5 without a second thought.
With Claude, it happens increasingly often that I ask a simple question and stare at the screen for 10 minutes wondering which one of us is having a stroke. This is NOT because of technical jargon — it’s specifically the uncanny case of “I know every word in this sentence, yet I have no clue what you were trying to say”.
I got a Kimi subscription for personal projects 2 months ago — switching frequently between Claude and Kimi made the contrast painfully obvious.
Kimi speaks in a way that is crystal-clear, compact, and it knows exactly HOW MUCH to say.
All three of these are severely underappreciated, but the last one is IMHO the hardest to get right. It’s pretty easy to find examples for what “speak clearly” looks like, but “which caveats are worth mentioning” is not a global knob but a million unique judgement calls for every topic.
Every time Kimi includes something extra in the answer, It’s a thing I genuinely needed to hear. As a very rough estimate, Kimi’s answers are ~60% as long as other models I’ve tried (in real sessions, not benchmarks) yet it doesn’t miss anything important.
It directly saves me time and effort when the model speaks like someone who’s goal is to convey the information in the most efficient way possible as opposed to sounding impressive or staying vague and abstract so you can’t call them out for being wrong because no concrete claim was ever made.
Please preserve this behavior
I find it hard to believe that you managed to achieve it without it being an explicit goal, but the lack of discussion about it concerns me. If somehow it was a lucky fluke, please try to figure out how it emerged so future models would not loose this invaluable feature.
The only thing preventing Kimi 2.7 from being a perfect coding partner is occasional endless looping. When it makes mistakes or gets confused, I can correct it — humans make mistakes too, and we manage just fine. The bottleneck is effective communication.