tips on saving token useage

[BackspaceOG]

Firstly, thank you to @Donchitos for making this. I am really enjoying seeing my ideas come to life.

Account: Claude Pro
Engine: Unity
OS: Windows 11 Home
IDE: Visual Studio Code

Q1) Token optimization strategies

I am finding that I'm burning through tokens fairly quickly, which fine, except I am about to hit my weekly limit. My game has 17 GDDs.

Things I am doing.

I have been using /clear often, but rn, Claude is now telling me that is nuclear option, and I need to just open a new session.

I am finding that VSCode does not print the CTX, so I am going back to windows terminal.

I have asked the agents to be more efficient, and it now uses --depth lean, but still likes to do design reviews often. In fact it just told me

"The real token drain on this project is the /design-review cycle. Each run spawns 4–5 specialist subagents, all
burning tokens. You've run it 14 times across 2 GDDs — that's the core cost."

It is also helpfully telling me

"Three things that would help most:

1. Use --depth lean from review 2 onward (you did this for some, not all — it cuts subagent count significantly)
2. /clear between GDDs — once Scene Management is approved, clear before starting Save/Load
3. Fix blockers in fewer passes — the input system took 6 reviews; aim to batch more fixes per revision before
   re-reviewing"

Is there anything else I should be doing?

I have some XP with development, but I am not a developer and certainly not a game designer. I am a product lead and this repo is made for ppl like me.

Q2) Account expectations

My game is a platformer, but it is pretty unique and has some interesting graphical needs. Is it unrealistic to expect to build with a pro account?

Thank you again to @Donchitos!

[ashmoonori-afk]

Brother i am here to ask same question on max 20.

[BackspaceOG]

I am already convinced I need to upgrade, worth it for what you get.

[Vandevious]

I also had this concern, so I configured Claude Code to run using DeepSeek V4 Pro - 85-95% as capable as Claude Opus for agentic coding, tool use, and narrative text but 5x-7x cheaper for input tokens and 7x cheaper on output tokens.

It's building a STEM Tutoring platform which gamifies First Principles and Socratic Teaching Methods while using Voxtral by Mistral to generate real-time voice guidance from the Chibli-styled robot mentor. It provided a lot of source material to build the First Principles and Socratic Teaching methods.

When I looked at my usage after completing the the 4th section of the 28 sections in the GDD - I was over 11million tokens (11M in, 70k out) used for $0.25 in usage.

[mahavira]

Same here. This is how many tokens have been used so far.

/architecture-decision Data Architecture: Resource-based DB + JSON Import

[BackspaceOG]

I also had this concern, so I configured Claude Code to run using DeepSeek V4 Pro - 85-95% as capable as Claude Opus for agentic coding, tool use, and narrative text but 5x-7x cheaper for input tokens and 7x cheaper on output tokens.

It's building a STEM Tutoring platform which gamifies First Principles and Socratic Teaching Methods while using Voxtral by Mistral to generate real-time voice guidance from the Chibli-styled robot mentor. It provided a lot of source material to build the First Principles and Socratic Teaching methods.

When I looked at my usage after completing the the 4th section of the 28 sections in the GDD - I was over 11million tokens (11M in, 70k out) used for $0.25 in usage.

This sounds like an epic project, how is it going?

I have come to realise that you need a max account for serious work, and I am not too serious at this stage and happy to chip away with pro for my part.

I have not looked into a config to get CC to use a different model. I will look it up, but I don't understand how that work, is it still CC at that point? I suppose CC is like a wrapper in your IDE, and the LLM it uses can be different to Opus or whatever model CC is using out of the box.

[Vandevious]

It's well worth it. I use the set of commands, below, to setup the model shift using Terminal/Powershell - right before I launch Claude Code. It only takes a couple seconds.

$env:ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
$env:ANTHROPIC_AUTH_TOKEN=""
$env:ANTHROPIC_MODEL="deepseek-v4-pro[1m]"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL="deepseek-v4-pro[1m]"
$env:ANTHROPIC_DEFAULT_SONNET_MODEL="deepseek-v4-pro[1m]"
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-v4-flash"
$env:CLAUDE_CODE_SUBAGENT_MODEL="deepseek-v4-flash"
$env:CLAUDE_CODE_EFFORT_LEVEL="max"

[BackspaceOG]

this is what Claude said, I was kind of right

That's a meaningful distinction: Claude Code's agentic design was built and tuned for Claude's specific strengths and behaviours. Results with other models can vary — some things work well, others (like very long multi-step reasoning chains or nuanced instruction following) may degrade in ways that aren't obvious until something goes wrong silently.
The cost arbitrage is real though. For teams doing high-volume routine coding tasks, it can make sense — just worth testing carefully on your actual workloads rather than assuming parity.

---

I am guessing deepseek would still do a pretty good job. However I am also guessing you would be paying for deepseek and CC.

Edit, I confirmed my understanding. Claude seemed to think the 48 agents would be an issue for DS, but ok for C# coding, not great for newer unity version I am using. will have a think and perhaps try, ty!

[Vandevious]

Nah, it's not an issue for DeepSeek, but check out the third-party benchmark testing for the models so you feel better about it. Deepseek V4 Pro within Claude Code and Composer 2.5 in Cursor perform VERY close to Claude Opus 4.7, in those same platforms/harnesses. The 5% improvement with Claude Opus 4.7 over would never be realized by the projects for nearly all users... while the real-world price per token cost would ABSOLUTELY be seen.

For instance, I have the $200 Ultra plan in Cursor (which is currenty giving you $400 in token use for the month). Using Claude Opus 4.7 I had already used 49% in the first week of May (so roughly $200 of the $400). At that point, I was fed up and go to a point where I was comfortable pivoting the context and codebase to Composer 2.5. In the next two weeks, I easily did 4x-5x the Agent use - pushed multiple large feature releases to my Veterinary Medical Record application and the token cost only went from 49% to 51% - ONLY 2% usage while on Cursor with Composer 2.5 as the AI model.

Check my previous reply, I provided the Powershell commands to use DeepSeek with Claude Code, instead of Opus.