Why my Claude Code token count exploded in 2026 — and what Anthropic changed

My public usage stats show something that looks suspicious at first glance: of 10.8B tokens processed with Claude Code, the overwhelming majority landed in the last three months. May 2026 alone was 3.9B tokens; June, 5.9B. Before April 2026 the bars are slivers.

I have used Claude Code daily since April 2025, and spent real money on Anthropic the whole time — €1,662 across 16 months of invoices, peaking at €180/mo from July to November 2025. So why does the telemetry look like I only just started?

Three reasons, and one of them is a genuine product shift worth writing down.

1. Anthropic massively raised Claude Code limits in May 2026

This is the headline. In a few weeks Anthropic roughly doubled what a paying user can run:

• 6 May 2026 — the 5-hour usage caps were doubled for Pro, Max, Team and seat-based Enterprise, and the peak-hour reductions were removed entirely.
• 13 May 2026 — weekly limits were raised by a further 50%, in effect through 13 July 2026.
• For Max specifically, the weekly Opus budget moved from ~50h to ~75h (Max 5×) and ~200h to ~300h (Max 20×).

The extra capacity came from Anthropic's compute agreement with SpaceX — the Colossus 1 data center, 300+ MW and 220,000+ GPUs.

The effect on my own numbers is unmistakable: the May→June jump lines up exactly with these changes. A higher ceiling simply lets me burn far more per week than was ever possible before.

2. "Tokens" is 97% cache reads

A token count is not a measure of effort. Of my 10.8B, 96.9% are cache-read tokens — the context Claude Code re-reads on every agentic turn. Actual input is 0.1%, output 0.5%. So the number tracks agentic intensity — long sessions, large contexts, thousands of tool calls — which is exactly what scaled up once the limits did.

3. The early years are simply missing

The deeper reason the chart skews recent: most of my 2025 telemetry is gone. A timestamp-less legacy history format, a July 2025 corruption event, and automated backups that only began in 2026. My spend from that era is documented; my tokens are not. The chart cannot show what was never saved.

And the pruning is not just history — it is live. Reconstructing daily spend from the backup shows ~$9.3k of API-equivalent usage over the last 90 days, but a local-only monitor (CodexBar reads only ~/.claude, which Claude Code prunes after 30 days) would see just ~$6.0k of it. The other ~$3.3k survives only because the dotclaude-sync backup stitched the pruned transcripts back together — the same reconstruction that now drives the daily-spend chart.

So the recency is real ramp + raised limits + an accounting quirk + lost history — not a late start. The full, sourced breakdown lives on the stats page.