Token economics: what 15 billion Claude tokens actually cost, and what they bought

Over one focused sprint of solo SaaS building with Claude Code I processed approximately 15 billion Claude tokens across the home Apple server — three Macs running Claude Code in rotation as one logical development environment. That number does most of its work as a headline. It's also worth pulling apart, because the breakdown tells you a lot about how agent-augmented engineering actually works — and whether it's worth the spend if you're considering the same workflow.

The breakdown

From the logged sessions across the home Apple server:

| Bucket | Tokens | Share |

|---|---|---|

| Cache reads | 11.6B | 77% |

| Cache creation | 378M | 2.5% |

| Output (model-generated) | 38M | 0.25% |

| Direct input (uncached) | 158K | <0.001% |

These totals are aggregated across the home Apple server — three Macs running Claude Code in rotation as a single development environment. 15B is the conservative rolling headline; the real total floats between 15 and 19 billion depending on the active sprint week.

What each bucket actually is

Cache reads (11.6B). This is the agent replaying recent context — the files in the active session, the running todo list, recent tool results. Every time Claude Code makes a tool call (read a file, run a grep, propose an edit) it re-reads the conversational state to stay coherent. Most of those tokens are the *same tokens* served back, billed at the Anthropic cache-read rate.

Cache creation (378M). New context being added to the cache as the session grows. Each new file Read, each new search result, each new system reminder bumps this up.

Output (38M). The actual model-generated tokens: code suggestions, edits, plans, reasoning, summaries. This is the bucket where "Claude wrote the code" literally lives.

Direct input (158K). Tokens that flowed in *without* a cache hit. Almost negligible — under one Pride and Prejudice's worth of text across the whole sprint of nonstop work. This is what most people imagine when they hear "input tokens." It's actually a rounding error.

Rough cost math

Anthropic's pricing for Claude Sonnet today, at Anthropic API rates:

Plug in the numbers above:

Total: about $7,000 across the home Apple server over the sprint.

I'm a Max plan subscriber — most of this is covered by a flat monthly fee. But the underlying compute economics are still worth understanding because they tell you what's possible.

What that bought

Concrete deliverables shipped in the same sprint:

Roughly 600K lines of code across 10+ active git repos, 580+ commits.

The conventional alternative — building this with a small engineering team — costs in the neighborhood of $200K per senior engineer per year fully loaded. Even at four engineers running flat-out for two months it's $130K in salary alone, not counting management overhead, hiring time, onboarding, and the inevitable team-shape friction. Tokens are not a tax. They're the cheapest part of the stack by an order of magnitude.

The operator lesson

The hiring conversations I'd want to have with someone evaluating this résumé go through one specific door: I have a *visceral* sense of what production AI costs and what it doesn't. I've seen the bill. I've watched the cache-hit rate drop by 12 points overnight when a system prompt was edited wrong. I've reduced p99 by 4 seconds by moving one section above the cache-control marker.

That's the difference between an engineer who *uses* AI and an engineer who *runs* AI in production. The latter has done the cost-modeling, owns the latency, and knows when to spend $0.05 on a Claude call versus when a deterministic regex would do the job for $0.

If you're shipping AI features at any scale, that intuition isn't optional. It's also one of the few things you can't fake on a résumé.

What 15B tokens doesn't tell you

A few honest disclaimers:

But the headline survives. 15 billion tokens is what a high-tempo solo SaaS sprint looks like in 2026. Anyone hiring AI engineers in a regulated vertical should be calibrated against that number — and asking candidates to show theirs.

More from BlueWave