Reallocating $100/Month Claude Code Spend to Zed and OpenRouter

1. The LLM provider doesn't know it's you (unless you have personally identifiable information in your queries). If N people are accessing GPT-5.x using OpenRouter, OpenAI can't distinguish the people. It doesn't know if 1 person made all those requests, or N.

2. The ability to ensure your traffic is routed only to providers that claim not to log your inputs (not even for security purposes): https://openrouter.ai/docs/guides/routing/provider-selection...

It's been forever since I played with LiteLLM. Can I get these with it?

If you're only using flagship model providers then openrouter's value add is a lot more limited

Not true in any non startup where there is an actual finance department

But if OpenRouter does better (even though it's the same sort of API layer) maybe it's worth it?

My coding is done with OpenCode with an OpenRouter API Key.

Going with KiloCode would be doing the same, but with some more layers.

And given that I can't see the on-demand API pricing, I'm really not convinced on how it would be an improvement.

They are blowing their "weirdness budget" on nonsense.

But their tab complete situation is abysmal, and Supermaven got macrophaged by Cursor

And I'm seeing the same thing in my sphere- everyone is bailing Anthropic the past few weeks. I figure that's why we're seeing more posts like this.

I hope they're paying attention.

It seems Cursor somehow builds a better contextual description of the workspace, so the model knows what I'm actually trying to achieve.

The problem is that with Cursor I'm paying per-token, so as GP suggested you can easily spend $100+ per month vs $20 on Claude Code.

Could it be related to this?: https://news.ycombinator.com/item?id=47660925

Do you think this would be a decent approach?

Also, which client would I use for this? OpenCode? I don't think Claude Code supports using other models. Thoughts?

Not everyone gets $1K of usage, and you don't know how fat the per-token margins are. It's like saying the local buffet place is losing money because you eat $100 worth of takeout for $30.

$1K is not actual cost, just API pricing being compared to subscription pricing. It is quite possible that API has a large operating margins, and say costs only $100 to deliver $1K worth of API credits.

It's obviously capital-subsidized and so I have zero expectation of that lasting, but it's pretty anti-competitive to Cursor and others that rely on API keys.

  My Claude Code usage stats after ~3 months of heavy use:

    Favorite model: Opus 4.6          Total tokens: 42.6m
    Sessions: 420                     Longest session: 10d 2h 13m
    Active days: 53/95                Longest streak: 16 days
    Most active day: Feb 9            Current streak: 4 days

    ~158x more tokens than Moby-Dick

  Monthly breakdown via claude-code-monitor (not sure how accurate this is):

    Month     Total Tokens     Cost (USD)
    2026-01     96,166,569       $112.66
    2026-02    340,158,917       $393.44
    2026-03  2,183,154,148     $3,794.51
    2026-04  1,832,917,712     $3,412.72
    ─────────────────────────────────────
    Total    4,452,397,346     $7,713.34

That’s what really appeals to me. I’ve been fighting Claude Code’s attempts to put everything in memory lately (which is fine for personal preferences), when I prefer the repo to contain all the actual knowledge and learnings. Made me realise how these micro-improvements could ultimately, some day, lead to lock-in.

> Run <other harness> in tmux and interrogate it how feature X works, then build me the equivalent as a pi extension.

I’ll give it a try!

Is OpenAI codex not also charging by usage instead of subscription when using pi?

I tested one of the other models that everyone is raving about yesterday (Qwen 3.6 plus) and within minutes found myself arguing with it even over a very simple task. After about 30 minutes (in which token usage never went over 50k because it was just me rewinding to give it more and more explicit instructions which it kept ignoring), I reverted everything and did it with Opus in literally about 4 minutes, after intentionally giving Opus a much more vague prompt.

Honestly, it seems like both of you were feeling a bit "grumpy" at the moment, but sending passive aggressiveness towards the maintainer you are trying to get to merge your code (or not your code, someone else's code?) seems like a very bold strategy regardless.

As someone else pointed out cooler heads and less passive aggressive responses would've resolved this issue easily.

I'm pretty conservative when it comes to clearing the context, and I also tend to provide the right files to work on (or at least the right starting point).

I had seen prior to using the model that it starts producing much worse results when the context used is larger, so my usage style probably helps getting better results. I work like this with Claude Code anyway, so it wasn't a big change.

Many of us got the annual Lite plan when they had the $28 discount. But even at $120 I think it's a good deal.

It turns it into a very good value for money, as far as I'm concerned.

GHCP at least is transparent about the pricing: hit enter on a prompt= one request. CC/Codex use some opaque quota scheme, where you never really know if a request will be 1,2,10% of your hourly max, let alone weekly max.

I've never seen much difference with context ostensibly being shorter in GHCP, all of the models (in any provider) lose the thread well before their window is full, and it seems that aggressive autocompaction is a pretty standard way to help with that, and CC/Codex do it frequently.

We are not allowed to use anything other than our company provided GHCP credentials due to the data retention clause in our contracts. Ie. they are not allowed to use our data.

The worst I saw - multiple parallel agents (opencode & pi-coding agents), with Kimi and glm, almost non stop development during the work day - 15-20% session consumption (I think it’s 2h bucket) max. Never hit the limit.

In contrast, 20$ Claude in the similar mode I consumed after just few hours of work.

For several models like Kimi and glm they have b300 and performance really good. At launch I got closer to 90-100 tps. Nowadays it’s around 60 tps stable across most models I used (utility models < 120B almost instant)

85% discount is actually a bit lower than I remember. I think it used to be closer to 90-95%. They're getting stingy ;)

But at that point we are just min/maxing the details, and all I can say is if you are on a $100/$200 a month subscription to any of these services and not using them regularly then you shouldn't be on a $200 subscription any more than you should be on a $700 a month gym membership when you go every 3 months for 15 minutes.

If I just let opencode zen run claude opus to plan and execute, I'd spend $20 in like 5 minutes lol

We are not the only one. I found other people online experiencing the same issue. It is hard to tell how wide-spread this is but it is strange to say the least.

The liabilities are completely offset by prepayments from your customers though. Even better, you can earn interest on the deposits without paying any out.

If you just dont want the liabilities on the books, issue refunds. Expiring credits feels like a cash grab.

So yes obviously you can do what you want as long as you abide by terms of service, but the terms of service does NOT allow you to resell the API.

OpenRouter is a valuable service but I’ll probably try to run my own router going forward.

https://openrouter.ai/docs/guides/overview/auth/byok

I use hooks to automate decisionmaking (i.e. stronger permissions than a regex by parsing the bash) and similarly automate guidance. Our tooling is open source so here is an example: https://github.com/Devleaps/agent-policies-server/blob/maste...

Another example is I whitelist dependencies based on the dependency age, for example a library needs to have been around for a year, then it can do `uv add {dependency}`.

As a third example, hooks tell Claude not to write out *_test|debug.py at the root of a project, which for some reason it very often does when it wants to fix some issue. The hooks tell it to write a proper test case using the test framework in place. So instead of having random debug and test scripts everywhere after a long session, I have more test coverage.

This is all in the agent-policies-server project linked above. Mainly it reduces interruptions and I don't have to worry about it doing something particularly stupid. (It is not a replacement for sandboxing)

> Do you use an editor alongside the CLI to review code or only examine the diffs?

I do have the file tree open alongside the CLI, and that is both in Zed. How much of the code I review depends on who owns the code, meaning a client, employer or me. In most cases I review it myself, as for many clients code goes through a peer review process. In some cases the organization uses automated quality metrics and has agents looking at code instead. If agents don't have any more comments, and the quality metrics also approve, it's good enough for them. As for my own personal projects, I look at the code when I feel like it, which is practically never.