The AI Spend Analyzer

Last updated August 12, 2026

This is a static site with no backend to upload to. Parsing, classification, and analysis all happen in your browser's JavaScript. Check the Network tab while you drop a file in and you'll see nothing sent anywhere.

This result is not saved. It disappears when you reload the page. Copy it as Markdown or download it before you leave.

Free tool

Drop your AI usage export in and see it for yourself

Works with an Anthropic Console cost/usage export or an OpenAI Usage Dashboard export — the parser auto-detects a few other shapes too. Everything below computes in your browser; nothing you drop in is sent anywhere.

Reading your file…

This is a static site with no backend to upload to — parsing and analysis run entirely in your browser's JavaScript. Check the Network tab while you use this tool and you'll see nothing sent anywhere.

What we look for

Nine finding rules, each computed from your own numbers. Nothing here is a guaranteed saving; it's what to check next, and each one links to the matching lever on Reduce AI Costs where one applies.

• Most of this spend sits on a frontier-tier model. That isn't automatically a problem; a genuinely hard task needs a frontier model. It's the single highest-leverage place to check whether every one of these requests needs that tier, since a routing change here moves more dollars than anywhere else in this file.
• Spend jumped hard from one day to the next. A sudden, large jump usually traces to a specific, findable cause: an agent loop that stopped terminating, a retry storm, or a feature that shipped without a rate limit. That's the runaway-agent signature CostMon's alerting is built to catch before it repeats.
• Spend has been climbing for several periods straight. A single spike gets noticed. A quiet, steady ramp usually doesn't, until it's a much bigger number than the last time anyone checked. Worth confirming this tracks a known driver, such as more users or a new feature, rather than a slow leak nobody flagged.
• A new line item started partway through this file. A brand-new charge is worth a quick check even when it's expected, such as a new model version or a feature that just shipped. It's the easiest place to catch a silent model upgrade, or a misconfigured deployment, before it compounds over a few more billing periods.
• Spend barely drops on weekends. A workload driven by people using a product usually falls off on Saturday and Sunday. Spend that keeps running at close to the weekday rate is the signature of a scheduled job, an agent loop, or automation nobody remembered was still on a timer.
• Some of this spend isn't plain token usage. Web search, code execution, and session-based line items bill differently from ordinary input/output tokens, and neither prompt caching nor model routing touches them. Check these against the provider's own pricing docs rather than the token-focused levers on this site.
• Many small, one-off lines add up. No single one of these is worth chasing alone, but a long tail of small, unrecognized experiments or one-off runs can add up to a real, recoverable amount that a glance at only the top line items would miss.
• Nothing in this file is broken out as cached or batch. That doesn't mean caching or the batch API aren't in use. A cost export grouped by model can't show cache hit rate at all, since a cache read still bills and labels as that same model's tokens. If you haven't checked whether prompt caching or batch processing fit this workload, it's worth a look regardless of what this file shows.
• Some labels didn't match a known model. Rather than drop these silently or guess at what they are, CostMon names them. They're still counted in the totals above; they just didn't match a pattern this tool recognizes as a specific model or provider.