Developer spends $8,857 on Claude Code across 6 projects in 14 days
Ethan documents spending $8,857.62 over 14 days using Claude Opus 4.8 across 6 simultaneous projects, sharing detailed cost breakdowns, token usage patterns, and workflow lessons about caching, model selection, and prompt discipline.
Score breakdown
The post provides concrete token-level billing data showing that cache management — not raw prompt length — is the dominant cost lever when using Claude Code at scale, with an 86.4% cache hit rate cutting what would otherwise be a far larger bill.
- 01Total spend: $8,857.62 over 14 days using Claude Opus 4.8 with a 1M context window
- 023.884 billion total tokens across 47,235 API requests
- 03Cache hit rate was 86.4%; cache hit tokens cost 1/10th of new input tokens
Ethan's Dev.to post offers one of the more granular first-hand accounts of heavy Claude Code usage, covering 14 days, $8,857.62 in API spend, 3.884 billion tokens, and 47,235 requests — all using Claude Opus 4.8. The six projects spanned a full-stack Next.js SaaS with payments and security hardening (LightCraft V2, ~$4,200), a multi-API video pipeline (~$1,800), a Playwright-based visa slot grabber (~$1,100), a psychology quiz app (~$600), a developer tutorial (~$500), and a simple API wrapper (~$400). Spending was not linear: the first three days cost ~$800 while Ethan was still learning the tool, two peak days hit $800+ each, and later days dropped to ~$300 as configuration improved and back-and-forth conversations decreased.
Of the 3.884 billion total tokens, 2.499 billion were cache hits and 1.322 billion were cache writes — meaning 86.4% of tokens were cached context reads at one-tenth the cost of new input.
The most counterintuitive finding concerns token distribution. Of the 3.884 billion total tokens, 2.499 billion were cache hits and 1.322 billion were cache writes — meaning 86.4% of tokens were cached context reads at one-tenth the cost of new input. Ethan argues this is why skilled users can achieve dramatically lower costs than careless ones: thrashing the context and invalidating the cache on every turn can multiply the bill several times over. On the output side, Opus 4.8 charges $75/M output tokens versus $15/M input, so verbose model responses with unnecessary explanations are a direct cost driver — Ethan's `CLAUDE.md` explicitly instructs Claude to be concise and skip unrequested explanations.
For model selection, Ethan uses a personal rule: Opus 4.8 for tasks that would take more than two hours of personal thinking time (architectural decisions, complex state management, edge cases), and Sonnet for grunt work like batch file changes, adding comments, or writing tests. He configures two profiles in `settings.json` and switches based on task complexity. The psychology quiz app — built in 4 hours from a single one-sentence prompt to a deployed app with Framer Motion transitions, radar chart visualization, and QR code sharing — is cited as an example of Claude Code's "sweet spot": clear requirements, well-defined scope, and polished output with minimal iteration.
Key facts
- 01Total spend: $8,857.62 over 14 days using Claude Opus 4.8 with a 1M context window
- 023.884 billion total tokens across 47,235 API requests
- 03Cache hit rate was 86.4%; cache hit tokens cost 1/10th of new input tokens
- 04Opus 4.8 pricing: $15/M input tokens, $75/M output tokens — output costs 5x more than input
- 05LightCraft V2 (full-stack video SaaS) was the most expensive project at ~$4,200
- 06Peak single day was June 18 at $828.74, running security hardening and video pipeline debugging in parallel
- 07Psychology quiz app went from zero to deployed in 4 hours from a single one-sentence prompt
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 20, 2026 · 08:55 UTC. How this works →