How GSD affects token consumption?

[Suwot]

I'm using Copilot in VSCode. There's basically nothing to complain, because my flow is highly mannual, and even though AI writes 95% of code, I'm still fullly invested in every line it writes.

It's working great but I want to keep track of new approaches, as this one stuck for me from day 1 as the only reliable way of coding with AI.

Before I try GSD I'd like to understand the downsides, like effect on token consumption. Can you guys share your estimations? I did find similar topic, but there's only 'high consumption' and 'it should help' kind of estimations, which is not the answer I'm looking for.

[trek-e]

GSD does consume more tokens than raw Claude Code, but with measurable structure benefits.

Rough estimate: 10–25% more tokens per session compared to unstructured prompting, primarily from:

• Planning artifacts loaded into context (.planning/ files)
• Structured agent prompts in skill files
• Checkpoint writing and STATE.md updates

What you get in return: The structured approach typically reduces total session count — you spend fewer total tokens on rework, backtracking, and context re-establishment across sessions. RTK (Rust Token Killer) is a companion CLI that compresses command output before it hits Claude's context window, which users report as saving 60–90% on tool result tokens.

For Copilot/VSCode: GSD works in any Claude Code environment. Token handling differs slightly — Claude.ai/Copilot both have context limits but the monitoring hooks work the same. Your manual review approach already sounds disciplined, so the overhead is mostly the planning artifacts which pay back in fewer corrections.

[Suwot]

thanks for detailed structured answer, I got the idea of what to expect

[RodrigoLoft]

rtk didn't seem to work with gsd, i am doing something wrong?

[x-iso]

well, I'm not so sure it's worth it with current obscene weekly limits. I have tried integrating GSD into my project, mapped codebase and everything, planned couple of tasks and then started executing. ended up burning through 40% of weekly limit in couple of sessions with this, and end result - half-done job on one task and complete mess up on second more important one. like, no research was done at all on how to implement it, no question asked either, it just seemed like it faked doing any job at all on this, with long-winded borderline bureaucracy stuff that didn't seem to do anything of use in the end. so far I'm quite disappointed.

[trek-e]

in our discord the discussion there a lot of people are dropping claude because of the limits. Codex has limits but doesn't seem to get hit as often. It appears the AI vendors are playing games at this point.

[PhenixH]

I join @x-iso in his evaluation. On Claude Max x20
I used GSD few months ago and everything was fine, on big project, I only hit once the 5 hours window limit (working on different subject in the same time), and last week, after only few steps i burned the full weekly limit (with less than 20% of the 5 hours window limit).
Something is different on how Anthropic count the usage when usage GSD (does it use the Agent SDK? Anthropic is officially changing how they measure usage with this approach).
Claude Update Plans concerning Agent SDK usage

[trek-e]

in my efforts to figure out ways to keep doing this, I've made some skills https://github.com/The-Artificer-of-Ciphers-LLC/skills-from-the-artificer specifically the teir routings. one thing I've noticed is that the higher level skills will read in a whole file looking for something, if it's a big file, that's expensive.

so I'm trying to force claude to use haiku or other models instead of opus with these. I juts finished and am starting using them to see. It may work, may do nothing, but that was my goal 'use cheap models to search for things, read/write files, find a line in a file' and then hand it off to the expensive model to do the work.

still an experiment but I put it there if you want to poke at it.