I Just Saw Someone Crowdsource Their Claude Code Bill – and the Numbers Are Wild

I have been running Claude Code almost daily since it dropped, and I have never had a clear picture of what my subscription is actually worth in API terms. The dashboard shows usage, but Anthropic does not translate that into "what you would have paid at platform rates." So when I stumbled on a community project that reverse-engineers exactly that, I stopped everything and read the whole dataset.

What the meter actually measures

The project is called Claude Code Meter, built by Chris Nighswonger at VSITS. It works by intercepting response headers from Claude Code sessions and benchmarking them against Anthropic's published API rates. The result is a clean dollar-for-dollar comparison between your subscription cost and what the same token volume would cost on platform.claude.com.

The current dataset is small but deeply observed: one contributor, 84,416 API calls across 237 sessions, tracked over 58 days from April 30 to June 2, 2026. That single account generated an API-equivalent value of $19,107 against just $193 in subscription spend. Yes, you read that right - nearly 100× return on the subscription dollar.

The tier breakdown

Here is the part that made me rethink which plan I am on. The meter scales the observed Max 5x account across all three tiers using Anthropic's 1× : 5× : 20× capacity ratio. The value multiplier is not what the marketing suggests.

Pro and Max 5x return identical value per dollar. The only real premium is Max 20x, which doubles your return and halves your cost per unit of capacity. If you are on Max 5x and using it heavily, the upgrade to 20x is mathematically a no-brainer.

Caching is the real multiplier

The plan tier is not even the biggest lever. The meter shows an 88% average cache hit rate on the observed account. At full API rate with zero caching, the same workload would cost $161,360. With the observed cache behavior, it drops to $19,107. Prompt structure and cache hits beat subscription tier by an entire order of magnitude.

This is the most actionable insight in the whole dataset. If you are not aggressively using cacheable prompts and reusing context windows, you are leaving money on the table regardless of which plan you pay for.

The Opus 4.7 mystery

There is an unconfirmed hypothesis floating in the data: Opus 4.7 may burn quota at roughly 2.4× the rate of 4.6. The catch is that the visible per-turn metrics actually show 4.7 as cheaper, not more expensive. The suspected reason is that adaptive thinking tokens are billed against quota but hidden from the API usage response. Until Anthropic exposes a per-visible-token quota field, this stays labeled as a hypothesis, not a finding.

Why this matters for the rest of us

The dataset is N=1. One contributor, one account, one continuous usage pattern. The magnitudes will not generalize until more people submit their headers. But the methodology is solid: every rate is verified against platform.claude.com/docs, and the raw data is open at /api/v1/dataset.

If you are a Claude Code subscriber, I would strongly consider adding your usage. The more data the community collects, the sooner we get a real picture of what "5× more usage" actually means, and whether the plan we are on is the plan we should be on.

I am planning to hook my own sessions into the meter this week. The GitHub repo is open source and the contribution path is just a few header exports. I will report back if my numbers look anything like the first dataset.

Source: meter.vsits.co - open data dashboard for Claude Code subscription analytics, maintained by Chris Nighswonger at VSITS.