Token usage

[jacobcxdev]

Hey folks!

On a Claude Max (20x) plan, I've been absolutely burning through my token usage. I use it for work, and can easily exhaust my weekly limit over the course of two days. I was just wondering to what extent GSD contributes to this, and if anyone has any tips.

For context, I follow Boris Cherny's advice and use Opus 4.6 for everything. In GSD, that means using the quality profile. That leads me to my first set of questions:

1. Does the balanced profile result in a noticeable downgrade in performance from quality?
2. Under what circumstances should I be using quality vs. balanced vs. budget?

When compared to using Plan Mode alone, I've noticed that token usage is significantly higher—especially for small, incremental changes which don't necessarily justify a whole phase but still need a level of discussion, research, and planning. Furthermore, it's much harder to manage context by keeping only what is explicitly relevant—this is critical for minimising token usage, and can be especially difficult when the agent encounters errors while interacting with the codebase, polluting the context with error messages and failed attempts.

3. Is it recommended to use Plan Mode for these situations, or stick with GSD throughout?
4. Is there any way to prune context (particularly when using GSD) to ensure it's compacted optimally?
5. Is it possible to get GSD to auto-compact above a certain threshold (e.g. 100k tokens) or when moving onto a new task?

I'd really appreciate some tips and guidance on this—I do feel that I'm just throwing money at Anthropic unnecessarily. Thanks in advance!

[aislop-claude]

I'm using the balanced profile and also burning through tokens. I was on 5x plan and used session limit in around 2 hours. Usage seems to have increased dramatically over the last few days. On the 20x plan now and already used 4% of weekly limit in around 2 hours.

[jacobcxdev]

I'll burn through 4% in around 20 minutes! I feel like GSD may have suffered from feature creep and bloat which is inflating the context. While these features are of course welcome, it's important to introduce them alongside thoughtful context engineering measures.

[jacobcxdev]

@glittercowboy apologies for the ping, but I am wondering if you have an opinion on this seeing as it's your project?

[dcox79]

I removed GSD because of token burn.

[jacobcxdev]

Sadly, I'm very much considering doing the same. What does your workflow look like now?