undefined | Better HN

0 pointskouteiheika13d ago0 comments

This. Here's a quick experiment I did yesterday.

I got a new $20 Claude subscription to try the new Fable model. I gave it a single prompt, and it barely finished, using up my whole session quota (it was at ~95% when it finished) and 10% of my weekly quota.

For comparison, with the Kimi Code $40 subscription I can pretty much constantly run two/three agents in parallel for the whole week, and I never run out of quota. I can blindly throw it at anything and everything without worrying about hitting the limits. (And it's not exactly a cheap model to run -- it has 1 trillion parameters!)

Is Kimi as good as Claude? Of course not. But you don't need the absolute state-of-art for most things. If I don't have exceptionally difficult tasks it makes no sense to use it. Just throw Kimi at it, and even if it needs to run 2 or 3 times longer in the background I don't care, because I'm not running out of tokens there.

0 comments

7 comments · 2 top-level

nl13d ago· 5 in thread

A word of caution on this.

I've tried this too, and was disappointed.

Kimi generally benchmarks at "a bit more intelligent than Sonnet Medium" levels[1] and I'd agree broadly with this assessment.

If you have adapted your coding to rely on the agentic style that is doable in Opus 4.7+ then you will find Kimi disappointing.

If you are using it in a more targeted way then it can work well.

[1] https://artificialanalysis.ai/agents/coding-agents?agents=cl...

kouteiheikaOP13d ago

Yes, I would agree with this.

I think it works best when you're using the agent in a more hands-on way with a targeted prompt. If you're obsessive about code quality like I am (so you thoroughly review and, when needed, reprompt or even rewrite what the agent does) then you'll be fine, but if you like to just throw a prompt at the wall and expect it to plan and execute the whole thing perfectly then you'll be disappointed.

A middle-ground trick one can use is to have Opus (or Fable now) plan the whole thing and get something cheaper like Kimi execute on it.

rented_mule13d ago

CodeWhale (formerly deepseek-tui) automates this over DeepSeek V4 Flash and Pro. My shallow understanding is that it prompts the model to evaluate the complexity of a given task, then decides on Flash vs. Pro at various reasoning levels for that task. This can help with both cost and speed. If other agent platforms don't already do this, I have to imagine they will at some point.

I'm retired and can't justify spending too much on these things. CodeWhale over DeepSeek is helping me understand this space much better (and have some fun!), and it's quite affordable. I've spent ~30 hours using it over the last couple of weeks, and I've spent $3.89 on DeepSeek in that time. If I don't feel like writing any code for a few weeks, I pay nothing. Looking at DeepSeek's dashboard, about 60% of my requests have gone to Pro and 40% to Flash. I've used 97M Pro tokens and 19M Flash tokens (well over 90% of each have been cache hits, so the price is much lower than it would otherwise be).

2 more replies

poly2it12d ago

Is there any open model that can emulate the agentic experience you get with Opus 4.7?

rstuart413312d ago

GLM 5.1 gets close to 4.6. It can happily run for hours and achieve a result. It given it bugs like a race condition that lead to a count being out by 1 after millions of operations, somewhere in a hundred thousand lines of C code littered with locks and atomic swaps, and it found (as did Opus). Most other models can't.

I'm using Fable now and GLM 5.1 doesn't really compare. But it's literally 1/20 the price. I can't use Fable for coding - it's too expensive. So now we have three levels of models - lightweight ones you dispatch en masse to find things, ones capable of agentic coding tasks that can run for hours like Opus, and GLM (and possibly open source ones - I've only tried a few), and now Fable, which is a truly helpful "architecture buddy". Fable still makes many, many, mistakes, so you have to review every word it writes.

nl12d ago

Not yet that I've tried, and I'm pretty systematic about test driving them.

I keep https://sql-benchmark.nicklothian.com/#all-data up-to-date with latest releases and try out most that score 24+.

GPT 5.5+ or Opus 4.6+ are the only things I find useful like this. Notably Gemini isn't useful in this way.

EagnaIonat12d ago

> This. Here's a quick experiment I did yesterday.

It's like running a sports car and then complaining it burns through petrol too fast.

The truth is the model while impressive is not needed for much of what people need.

Local models can do the work and just offload heavy lifting to the cloud models.

j / k navigate · click thread line to collapse