Loading page
About this tool
Analyze AI usage logs locally with user-supplied token rates, model cost summaries, and counts-only personal-data or secret signals.
LLM Cost & Privacy Analyzer reads an AI usage log — JSON, JSONL, CSV or TSV — in the browser, totals input and output tokens per model, and prices them with rate tables you supply yourself, using cost = (tokens ÷ 1,000,000) × your rate per million. Alongside the cost breakdown it scans message and prompt fields for personal-data patterns such as email addresses, phone numbers, payment cards, IP addresses and API keys, and reports only counts per category — never the matched values. It is built for the engineer or finance owner who has an export of API usage and needs a spend figure and a privacy sanity check without sending that log to a third party.
Open LLM Cost & Privacy Analyzer on AltFTool — it loads instantly in your browser.
Drop a usage log on the Upload File dropzone or switch to the Paste Log Text tab — .json, .jsonl, .csv, .tsv and .txt are accepted up to 8 MB.
In Token Pricing Rates, replace the example presets with your own contracted rates per million tokens; a "*" entry prices any model not named.
Read the per-model cost table and the counts-only privacy findings, then take a JSON Summary, CSV Report or Markdown Summary that excludes prompt content and model names.
Costs come only from rate tables you enter, so the numbers reflect your contract rather than a list price that changed last quarter.
It looks for usage.input_tokens, prompt_tokens, inputTokens and a dozen other spellings, and flattens nested records up to five levels deep.
The scan reports how many matches of each category it saw; the exported report deliberately excludes model names, log content and the matched strings.
Each record's cost is (input tokens ÷ 1,000,000) × your input rate plus (output tokens ÷ 1,000,000) × your output rate. Rates come from a JSON table you paste, with a "*" entry acting as the fallback for any model not listed by name; a record with tokens it cannot allocate to input or output is left unpriced rather than guessed.
Up to 8 MiB of source text and 5,000 records, with the privacy scan reading the first 500,000 characters of content fields. If either limit is hit the results are marked as truncated so you know the totals cover only part of the file.
Content fields — content, prompt, message, input, text, requestBody and responseBody — are pattern-matched for categories including email addresses, phone numbers, payment cards, bank and IBAN numbers, national IDs, IP and MAC addresses, and keys or tokens. Pattern matching finds likely instances, not all of them, so a zero count is not proof a log is clean.
No — the preset buttons only pre-fill the rate table with example figures so you can see the format. Provider pricing changes often, so replace them with your own contracted rates per million tokens before trusting any total.