How accurate is the AI cost estimate?

When you provide real usage (a CSV or your provider's usage API response), the model-level cost is computed directly from token counts and a maintained rate card, so it is close to your actual bill minus any custom discounts. The website mode is assumption-based: it can only see client-side AI usage, not server-side API calls, so it gives a ballpark from detected features and the volume you enter. Every number shows the rate and assumptions it used.

Is my usage data uploaded anywhere?

No. CSV parsing and the Connect-mode JSON both run entirely in your browser — nothing is uploaded or stored. The only server request is the optional website scan, which fetches the target site you name, not any of your data. Your provider API key, in Connect mode, never leaves your machine: you run the usage call yourself and paste the result.

What savings does it suggest?

Right-sizing frontier models to a cheaper sibling for simple tasks, prompt caching for repeated system context, moving non-realtime jobs to the Batch API (about half price), trimming output tokens on verbose responses, and provider arbitrage on frontier spend. Each recommendation shows a dollar-per-month impact, an effort level and the assumption behind it, sorted by impact.

Which providers and models does it cover?

The built-in rate card covers OpenAI (GPT-4o/4.1/o-series/3.5), Anthropic (Claude Opus/Sonnet/Haiku), Google Gemini, Meta Llama (hosted), and Mistral, including cached-input and batch tiers. Unknown models are flagged and excluded from computed cost. Rates are editable estimates — verify against your provider's current pricing.

Can it read my ChatGPT or Claude usage directly?

For the API platforms and Team/Enterprise admin accounts, yes — OpenAI and Anthropic expose usage/cost APIs whose JSON you can paste into Connect mode. Consumer ChatGPT and personal Claude plans have no usage API, so for those you would use the CSV export where available.