OpenRouter (Provider only)

[crimson flower]

https://docs.openclaw.ai/providers/openrouter

[somber tinsel]

I've spent $27 with 1k requests and 84.6M Tokens via OpenRouter and Kimi K2.5.

[little condor]

the new hunter alpha stealth model is amazing for openclaw, been using all day

[prime kayak]

What have you been using this on primarily? What’s the cost like?

[azure compass]

what do people think it is?

[little condor]

its free for now

[little condor]

no idea, if you ask the model itself sometimes it says claude, sometimes deepseek, no one knows but the way it talks about china is probably a chinese model

[spiral fable]

damn nice. didnt even see this. didnt realize it was free. how do you find it? any tool calling issues?

[glad hound]

When I use openrouter on free models I get an upstream inference error and it breaks. I have been avoiding using openrouter but think it might be time to look at why this is happening. Anyone have similar issues and resolved?

[silent osprey]

Good to hear it's working well in openclaw. I've been using it as player in a Mafia game, and it was incredibly bad at following instructions. Half of the time, it couldn't get a turn to speak because it failed to state an urgency when prompted and tried to speak out of turn instead. Even the tiny 7B model I use for local testing gets that part right...
Although the gameplay itself was solid. Good discussion, and Town correctly voted out both Mafia on day 1 and day 2---in a game without town roles (not implemented yet), that's impressive.

And on more general terms: In early February, we had GLM-5, and mid-February, we got one that hasn't been disclosed but, in my opinion, probably was Mercury 2. So this one could be anything, but that there are two (Hunter and Healer) limits the candidates. Not many companies train their models in "small/big" pairs at the same time.

[shy storm]

Hunter fails at visual tasks, it seems, and it totally hallucinates the results. I had it try to pull contact info from a business card and insert into crm, as I've had gpt 5.2 and 5.4 do without issue. Unfortunately it seems this model is lacking in that area.

[zinc herald]

I got this error when i hit the monthly quota I had set (check under Settings > API 'Limit')

[glad hound]

Thank you. I’ll check it out. I believe I set it low when I was testing and never thought to check back

[silent osprey]

They have. But you need to set custom parameters. https://openrouter.ai/docs/guides/routing/provider-selection The title is "provider selection", but it can also do models.

[zinc herald]

👍 I heard - though I haven’t confirmed it - that if you have less than $10 ~$20 in credits then you will get throttled down to lower speed.!

[glad hound]

I did see that you get more free ceiling if you put on $20 so I did. Llmapi is decent though. I primary codex as I have a workspace (many). Then use llmapi for fallback. Works well.

[silent osprey]

https://openrouter.ai/docs/api/reference/limits

[somber tinsel]

What is everyone using now instead of hunter-alpha? It's now Xiaomi: MiMo-V2-Pro and I'm super sad because I really really liked Hunter-Alpha.

[agile tartan]

appearently its free first week still:
📢 Xiaomi's all-new MiMo series models MiMo-V2-Pro, MiMo-V2-Omni, and MiMo-V2-TTS are now officially launched. Free globally for the first week in partnership with multiple mainstream development tools! ✨

but you probably need to configure api to go through xiaomi directly?
https://platform.xiaomimimo.com/#/docs/pricing

[dusky belfry]

wait what? that's expensive. I'm using AIsa and I only spent bout $10 or less with over 1k requests. I'm on kimi 2.5 ... Looks like AIsa is somehow cheaper then. Same thing for Qwen and even Claude Opus and the rest

[cold oriole]

im using kimi too but idk why he can lying and cant just do my basic request lol idk whats wrong with him, wbu?

[dusky belfry]

Works pretty well for me, fr

used it in my setup and it worked, tried it on AIsaClaw and it worked really good fr. was able to do most of my tasks

Maybe this is an issue with your provider?

[silent osprey]

Does anyone have a good strategy to deal with those ... "tricky pricing" providers that doesn't involve me checking the providers every couple of days and either blocking them manually or manually allowing good ones?

It feels like there's a new one popping up that undercuts the rest, but is effectively the most expensive one because they have no cache pricing, every time I check provider listings... :(

Do those even have their own infrastructure, or are they proxying another provider that has cache pricing and making their money that way? I mean, the lowest effective price is $0.321 in this case, that's a great margin...

[steel vortex]

Qwen3.6-Plus is free right now on OpenRouter. It has pretty good reting on ClawEval https://github.com/explaindio/ClawEval

[dusky belfry]

check other providers and their pricings and compare with this. Check AIsa and use it to compare

[normal quail]

If you want to refill from wallet on openrouter to get 1000 free tier prompts for 10.50 usdc, take it out in the Base network. And beware of gas fees, especially bad is ethereum ERC20. Avoid that

[crude thunder]

I did use stepfun until now bc the quality was well enough and it´s pretty cheap.
Would the jump to Qwen3.6 e.g make a huge difference?

Also thought about going from openrouter to the 20€ OLLMA subscribtion

[balmy hollow]

On the Openrouter page for a model it says openclaw is one of the providers for that model but when I run onboard it’s not in the list — is this normal?

[steel vortex]

I spend $20/month on Ollama and $10/month on Alibaba, which serves as my fallback. My main setup uses KIMI K2.5 for primary agents and most sub-agents, while I use MiniMax 2.7 for some sub-agents mainly because it’s faster. Ollama is the only one that offers MiniMax 2.7; Alibaba only has the older MiniMax 2.5, so when I run out of Ollama quota, I switch to KIMI K2.5 on Alibaba. I also run Qwen 3.5 locally for some agents. For staying on budget, I think a hybrid cloud + local approach works best.

[crude thunder]

sadly, I have no hardware to keep the models running locally :/

I guess I´ll be going with Codex for coding tasks and Ollama for the agents other tasks + OpenRouter as fallback

thank you for your insights 😄