I burned $13,836 worth of Claude Code tokens in June. The same work on open models: $3,106. That's $10,730 left on the table — one month, one laptop. So I built the instrument that shows you the gap. Open source, 100% local. Launching today 🧵

Frontier pricing is going up, not down. Fable 5: $10/$50 per Mtok — exactly 2× Opus 4.8, 5× Sonnet 5. Opus 4.8 fast mode? Also $10/$50. And agentic workflows multiply the burn: one prompt fans out into dozens of subagents. Coding factories are incredible. So is the bill.

Today I'm launching Caliper. See every dollar of your Claude Code spend — per session, per workflow, per subagent, per tool call. It reads the transcripts Claude Code already writes — right on your machine. No account, no telemetry. Open source, MIT.

The estimates are cache-aware and parity-tested against ccusage. What surprised me personally from my own spend? · the main chat is often the biggest line item — 41% of one $622 session (I need cheaper workflows!) You can't fix what you can't see.

It doesn't just report — it closes the loop. One click copies your real numbers + live API pointers into a prompt. Claude reads its own spending and (with your consent) writes you a personalized cost-discipline skill. /optimize-spend ships free in the plugin.

Try it now in Claude Code: /plugin marketplace add Cost-Caliper/caliper /plugin install caliper@caliper Then run /caliper — the dashboard opens on your current session. Free. Alpha. MIT.

Step 1 of 3 — See it. Route it. Own it. Next: smart routing to open models (GLM 5.2, DeepSeek, Qwen, Llama — American-hosted). Then: custom models fine-tuned on how you actually work — that take over once they beat the frontier model your agent would reach for.

The dashboard + /optimize-spend skill are free today. The waitlist gets you routing + training early access: → What did you burn last month? Run /caliper and reply with your number 👇

Github:

@miroyato @0x6b8

Nice work. We should collab. Been working on a solution to the same problem that uses multiple optimizations. Also open source and also locally run. Check it out at @tokenjamdev

This is sick! How does the optimization skill work? Is that just claude coming up with its own heuristics for how to cut token spend?

Right so the optimization skill first and foremost surfaces Claude's data to it, so it can actually review how it's working and create more optimized usage plans IE: workflows where work is delegated in a smarter way, or with subagents. Later we'll help users with a hosted solution for this.

Interesting I'm not super familiar with Claude's plugin architecture - possible to translate this over to Hermes setups easily?

@arjunbhuptani That’s a great question- I will look into it!

Wow, 10730 lost! Think?

Yeah totally avoidable too.i need to tune my workflows better

Very helpful - thanks Dennison

这差距看着是真有点肉疼

So real

$13k pain is real saw my own bleed this month. caliper in the plugin store? game changer

I saw it and was shocked. I'm using a bunch of subscriptions right now which cuts the pain a bit, but the reality is those subsidies are going away and then I'm going to have to fend for myself with costs.

Started deep diving into my usage and see so many opportunities for cost and time savings with some harness tweaks

It's pretty wild how many oppertunities there are to improve your spending. Especially since, for example Sonnet 5 is actually more expensive than Opus on High settings (but worse quality). It means you really have to tune in your spend or go broke.

@eigencloud Shut up and take your vibe coded shit elsewhere

Also the opportunity to optimize here is incredible.

Need it

Literally i could be saving like 70% using teams like @FireworksAI_HQ and GLM-5.2 Next step would be to benchmark again them to see what flows could be substituted directly.

Fiiirrreee

I'd love your feedback in particular on it!

You should definitely try in that case.

can't wait to try this!

Thanks @blake41!

This is the kind of measurement I wish more coding-agent demos showed. Do you break the saved cost down per task, per tool call, or per failed/rolled-back run? That feels like where a student builder could learn the most.

So many folks I know need this!

Awesome project. Looking forward to giving it a try. Thanks for open-sourcing it!

Thanks! Give it a try!

With many company employees on endless api plans, being mindful of spend seams like the only way to keep access sustainable. Personally, I always start with a cheap model (e.g. DS4 flash) and only escalate to more expensive models when I'm not happy with the overall quality or hit a wall.

Yeah especially now that subsidies are going to be ending we're going to need to figure out what our actual spend looks like.

killer

Thanks @tommylower ! And thanks for catching the signup bug, already fixed!

13k? rookie numbers

@arjunbhuptani And that was with fable suspended half the time!

Well explained. 📚

Facts!