Detailed Analysis
A developer's discovery that a $20/month Claude subscription generated $7,077 worth of API-equivalent computing work over five months has surfaced a growing tension in how AI companies price and package their products. The finding, shared on Reddit alongside a new open-source CLI tool called "burned," reveals that a single user's Claude Code sessions consumed 8.9 billion tokens across 323 sessions and 47,467 requests — value that would cost hundreds of times more if purchased through Anthropic's standard API pricing rather than through a flat-rate subscription plan. Perhaps more striking is the claim that a $200/month Max plan translates to roughly $1,400/month of pay-as-you-go equivalent usage, a 7x multiplier that illustrates just how aggressively subscription tiers are being priced relative to metered API costs.
The tool itself addresses a real gap in AI tooling: opacity around resource consumption. Claude Code, like many agentic coding assistants, stores detailed session logs locally in a directory structure (~/.claude/projects), but Anthropic provides no built-in dashboard translating raw token counts into cost-equivalent terms or breaking down spend by project, session, or model. The developer built a lightweight, read-only Go binary that parses these logs entirely on-device — a deliberate design choice ensuring no code or prompts leave the user's machine, which matters given that developers often work with proprietary or sensitive codebases through these tools. Notably, the tool also parses OpenAI's Codex CLI logs, surfacing an additional $1,600 of usage the developer hadn't previously tracked, suggesting that the visibility problem spans the entire agentic coding tool ecosystem rather than being unique to Anthropic.
The data also highlights the economic mechanics underpinning modern LLM pricing: prompt caching. The user's logs show caching saved $31,293 against a theoretical uncached cost of $38,370 — meaning caching accounted for roughly 80% of the theoretical spend and dwarfed the amount actually "burned." This underscores how heavily both providers and users now depend on caching architectures to make repeated, iterative agentic workflows (where large contexts, codebases, and conversation histories get re-sent constantly) economically viable at scale. It also explains why a single runaway session — in this case, an overnight autonomous agent loop that racked up $776 in equivalent value — can become a meaningful cost event even under a subscription model with no visible dollar signs attached.
This anecdote fits into a broader pattern across the AI industry in 2025-2026: as coding agents like Claude Code, Codex CLI, and similar tools move from occasional assistants to long-running autonomous workers capable of looping through tasks unsupervised, the gap between flat-fee subscription pricing and actual compute consumption is widening rapidly. Companies like Anthropic are absorbing substantial infrastructure costs to keep power users on subscription tiers, a strategy likely aimed at driving adoption and lock-in during a competitive land-grab phase before usage-based pricing or tighter rate limits become necessary. For end users, tools like "burned" represent an emerging category of third-party observability infrastructure filling gaps that AI vendors have left unaddressed — mirroring how cloud cost-management tools (e.g., FinOps platforms) emerged once cloud computing bills became unpredictable and opaque. As agentic AI usage scales, expect more grassroots tooling, and eventually first-party dashboards, aimed at demystifying exactly where compute — and money — is going.
Read original article →