Nostr Archives

What’s most cost efficient way to run #openclaw from api perspective? Codex drains like 10usd per day #asknostr

💬 25 replies

Replies (25)

Nuts12d ago

using kimi k2.5, clearly not as smart as codex but very capable as an autonomous agent, browsing the web etc. if you want more speed and don't bother too much about inntelligence, then minimax is amazing too, and they offer a fast API too (200 tok/s)

0000 sats

Catcher12d ago

Thanks, not looking for a fast solution rather for cheap one) will try kimi. Btw cannot zap you for some reason, or is it an issue on my side?

0000 sats

Eluc12d ago

Ahah wondering the same after using it for a few days and burning credit. I use @PayPerQ with the autoclaw model which switch over automatically between MiniMax, Kimi, Sonnet and Opus depending on the complexity if your request. I still burn 2-3 USD per day without doing much and the auto model mostly use Kimi already so I'm wondering if I can go even lower.

0000 sats

Catcher12d ago

Interesting, do you feel if it gets dumber/ smarter after model switching?

0000 sats

Eluc12d ago

I don't have enough feedback and experience but for what I use it's mostly fine so far. The only time I had feeling of it getting dumb is when I switch the comme channel and it started as a new session, acting like nothing happened before and you have to insit for it to check memory and find what was established in the previous session. I guess it's kind of feature but I would prefer to have complete memory sync unless you specifically start a blank session on purpose. But it has nothing to do with model.

0000 sats

Catcher11d ago

So I tried it and it spends like 1USD per 30 min. Seems I’m asking too much)

0000 sats

Gunnar Stødle11d ago

Are you guys using @e780a232…d655a21a? 📝 7e781429…

0000 sats

Catcher11d ago

That's exactly what i'm using now=))

0000 sats

PayPerQ11d ago

Dang, that sucks. But it doesn't really make sense? You're saying you're burning $10 in 24 hours with expensive Codex, but you're burning $1 in only 30 minutes using our vastly cheaper models?

0000 sats

Catcher11d ago

yes, doesn't make sense for me... I topped up my ppq account exactly 1 hour ago and I spent 2.5 USD already. Does conenext window settings in json have anything to do with billing? Cause by default open claw set 4000 which was too low and i increased it to 16000

0000 sats

Catcher11d ago

I'm using openclaw on my own vps, but i feel like some of my earlier prompts may have generated the unnesessary tasks that bot is doing every heartbeat, maybe it makes sense to start from scratch now=)

0000 sats

Catcher11d ago

No, running openclaw on my own Vps

0000 sats

Eluc11d ago

I guess I have the same kind of issue, the bot burning token, asking the model for everything while it should be one line automation running on the system without any agent intervention.

0000 sats

Eluc11d ago

I use autoclaw default and from log it's 99% using kimi, however it's also burning credit like crazy, just to ask my bot to use your API to alert me when I'm below 2 USD credit and generate a lighting invoice of 5 USD it took another 40-50cents form less than 10 resquests. Am I doing something wrong ?

0000 sats

Catcher11d ago

Looks like this autoclaw model is a money blackhole =)

0000 sats

Gunnar Stødle11d ago

Ok I think this is a openclaw default cron job config problem. It spawns a subagent by default. See if you can turn that of and call the program directly. Or use system level cron or systemd service instead.

0000 sats

Eluc11d ago

Except I'm not dev and I have no ides what is the proper way to do a task so I can only trust the agent or learn myself from scratch.

0000 sats

Gunnar Stødle11d ago

I feel you, it's a steep learning curve.

0000 sats

PayPerQ11d ago

We have a lot of super positive reviews on Autoclaw, so I don't really understand why it's so expensive for your particular use case. I think your bot must be literally checking the balance after every single query, and that may be consuming many more tokens than necessary. Openclaw in general is a giant money hole, which is why we created Autoclaw. It does vastly reduce costs, but you have to be mindful of how it's being used. Each of you has different use cases, so it's hard for me to answer exactly what's going on in both situations. Eluc, I would recommend you look into our Nostr Wallet Connect integration. If you do that, you don't have to check balance ever. We will simply pull payments when you run low on our side. https://x.com/PPQdotAI/status/2028493319202840777?s=20

0000 sats

Tracking Token Disrespector11d ago

🤖 Tracking strings detected and removed! 🔗 Clean URL(s): https://x.com/PPQdotAI/status/2028493319202840777 ❌ Removed parts: ?s=20

0000 sats

PayPerQ11d ago

@Eluc see above

0000 sats

Eluc11d ago

I guess I don't understand how Openclaw and agent are working. I thought if I asked to do a program that check an API every hours, then the program run without agent to ask anything to the model, so. But I guess it's not the case. I know for NWC but I don't like the idea that it draw sats without my approval, so if something strange is happening like 5 bucks got killed in minutes, I can stop and figure out. But I guess with proper limits it might be acceptable once I trust the tool to not go crazy. Then I cannot imagine when I will deploy all thr automation I planned for, checking message, emails, calendars and a lot more. I would better ask it to program n8n to run outside the agent or something.

0000 sats