Post by @decryption@aus.social

i have been mcuking around with opencode and openrouter for some llm fuckery because I wanna see what a more advanced "harness" can do and goddamn I burned through $20 of tokens so quickly using Kimi K3 - would have lasted like 5 minutes with GPT Sol or Fable

it shows how subsidised something like Claude or ChatGPT is compared to the raw API token costs for the same output

like if there's a future where someone vibecodes an app to suit them instead of using something off the shelf - the off the shelf app will likely be cheaper!

16 Replies

0 Boosts

Viraptor

@viraptor@cyberplace.social 4 minutes ago

@decryption
If we're talking long-term, we're still up for some architecture changes. Currently the labs can (still somehow...) burn piles of money so they do. Once they feel the squeeze, we'll see some serious experimentation in what's possible. (Where's my small coding model which hasn't been trained on Harry Potter fanfic? Where's the near-linear recurrent version?)
That will hopefully affect the inference energy use and pricing.

Matt

@mattdarveniza@mastodon.social 2 hours ago

@decryption I'm not even tokenmaxxing that hard at work and:

Matt

@mattdarveniza@mastodon.social 1 hour ago

@decryption ha no that's 23 days. Averaging about $100 per day right now with Fable.

josh's harp

@Joshsharp@aus.social 1 hour ago

@mattdarveniza @decryption I have a friend who's going all-out on this stuff and apparently hitting a $2k/week spend limit regularly 😐

Matt

@mattdarveniza@mastodon.social 1 hour ago

@Joshsharp @decryption for me the bottleneck is still my brain. It can pump out code and words all it likes but I still need to read it all and actually understand it. My assumption is that people burning tokens like crazy aren't actually taking that time.

decryption

@decryption@aus.social 1 hour ago

@mattdarveniza is that just today's usage???

minyak

@minyak@aus.social 2 hours ago

@decryption i can't get myself to pay for more than deepseek v4 flash.

also, exercise is great. You were carrying around another 50kg of body weight, so squatting a 20kg barbell should be… not onerous…

So good: improved sleep, improved travel (lifting bags etc), look better in the mirror. Everything is better.

50kg man: that's an achievement. Good work.

ǝlqɯnɹ uoɯᴉS

@shermozle@aus.social 2 hours ago

@decryption Kimi has a coding plan which is pretty good. K3 is a token pig. K2.6 is more efficient but not as capable.

GLM5.2 from z.ai is the sweet spot there.

I couldn't get OpenRouter to reliably use the coding plan though but you can configure both in OpenCode.

If you want a lighter harness, pi is good too.

decryption

@decryption@aus.social 2 hours ago

@shermozle oh yeah, im just fucking around seeing what can be done with a few prompts - know your enemy etc

ǝlqɯnɹ uoɯᴉS

@shermozle@aus.social 2 hours ago

@decryption try asking about what happened in Tiananmen Square in 1989.

Peter Evans

@evanspw@aus.social 2 hours ago

@decryption The Tokenicide is upon us.

decryption

@decryption@aus.social 2 hours ago

@evanspw $150k for a meatsack or $150k in tokens - the meatsack with a compsci degree will work for 1,000hrs/yr, i don't think you'd get more than 1000hrs out of a frontier model for $150k right now

but you have to deal with the squishy meatsack's "feelings" and "opinions" instead of bossing it around like a robot

Peter Evans

@evanspw@aus.social 2 hours ago

@decryption Meatsack good for 2000 hrs/yr (out of 8760 to choose from). Model costs are going to go up. Old mate J Huang is already telling people data centres will be about twice as expensive as they think (motivated reasoning though!)

decryption

@decryption@aus.social 2 hours ago

@evanspw i figure if someone's got a magical way to reduce token usage, they'd be frothing to implement it - but that doesn't seem to be the case

1 more replies