(posting to tech support but pls let me know if more appropriate elsewhere).
I’m on the Claude Max 20x plan and I’ve hit a brick wall. Since yesterday, my account is showing I’ve reached my Weekly Limit, even though my 5-hour session window is clear and I’ve barely used the web UI manually this week (and my Rebel usage has been very low compared to usual as I am on holiday).
It’s clear that Mindstone Rebel is consuming my subscription quota in the background much faster than expected. I suspect this is tied to the April 4th policy shift regarding third-party agent usage on personal subscriptions.
Before I consider switching to a pay-as-you-go API (which I know will be way more expensive), I’m looking for advice on how to optimize Rebel to play nice with the Max 20x limits:
Rebel-Specific Optimization: Does anyone have tips on reducing the “token burn” within Mindstone Rebel?
Are there settings to limit the “Search Depth” or the number of autonomous loops it runs per task?
Has anyone found a way to stop it from re-sending the entire project codebase/history with every single “thought” step?
Model Efficiency: Are you guys forcing the agent to use Sonnet 4.6? If so, have you noticed a significant “life extension” on your weekly quota compared to using Opus?
Local Hybrid Setup: Is anyone offloading the “reasoning loops” to a local model (like Llama 4 or Qwen 3) and only using the Claude subscription key for the final output? If so, how are you routing that in the Rebel config?
Prompt Caching: Does Rebel support Anthropic’s Prompt Caching? I’m wondering if my limit is being hit because the agent isn’t properly utilizing cache breakpoints, causing me to pay “full price” for the same context over and over.
Being locked out of my entire Claude account by Tuesday is a nightmare. If you’ve found the “sweet spot” for Rebel settings that stays under the Max 20x weekly radar, please let me know!
Thanks for raising this - Anthropic has recently started to reduce Max limits and made it harder to use Max subscriptions outside of Anthropic products at the same time.
We are very aware and working on various solutions, but my personal set-up at this point actually uses Opus as the main model and ChatGPT’s GPT-5.4 as the “worker” model. This way I use my ChatGPT subscription for most execution, while Opus still orchestrates what’s going on.
Happy to help you get this set up if it’s unclear?
Thanks Josh - I might give that a go as well while you guys work on other solutions… would you mind clarifying - when you say you use your chatGPT subscription for the “worker” model / “most execution” do you mean you use the pay-as-you-go OpenAI API key? Or is there a way to use a ChatGPT pro or another subscription for ChatGPT all-you-can-eat token usage for Rebel?
… also - is there a reason you are using ChatGPT 5.4 rather than Sonnet as the “worker” model? (as using Sonnet would still keep within limits with Claude Max I guess?)
I use my chatgpt account, which you should be able to authenticate in “agents & voice” in settings. I’m afraid Anthropic is quickly phasing out the ability for people to use their Claude accounts in anything but claude, so we’re actively working on making this as simple as possible, with a few options live next week I’d imagine - but this one is already a good starting point
Hi Josh, I set up my OpenAI account, but I’m seeing weird behaviour with automations. What I’d reeeeeeally like would be to be able to use Anthropic models through Bedrock, since my company mainly uses that. Is it already possible?
Hey there, that’s totally possible. You can set up a “custom profile” with your anthropic bedrock API endpoint + api key and you should be good to go? After setting them up as profiles, you just need to set them as “thinking model” and “working model”. 1 profile per model.
Could also set this up as a custom provider (bedrock) and then have the various models set up as profiles
Hi @Joshua so I set up OpenAI on Rebel using my ChatGPT Pro subscription as you have done in your setup - it is showing as “ChatGPT subscription active” - however when I use Rebel it is depleting my OpenAI token credits (I can see usage $ going up in platform.openai.com) - any idea what is going on? I haven’t given an API key to Rebel for OpenAI, just gone through the chatGPT login process…
FYI - yesterday I found that using the Claude API key, I managed to spend $65 in tokens just on routine tasks (scanning emails and what’s apps, importing call summaries) in an hour or two… I’m worried that the costs of running Rebel are going to balloon significantly
…. I see in the “Usage” tab within “Settings” that usage under Claude subscriptions are mapped separately, but it doesn’t seem to separate out usage under ChatGPT subscriptions…
After the update, I had to remove my OpenAI API key and re-log-on to ChatGPT to “force” Rebel to use my ChatGPT subscription, but it’s now working. Question - can we use the ChatGPT subscription to also power voice?