Under the hood
What a reading actually costs
Most of this tool is free by design, and the one part that is not costs about a cent. Here is the whole pipeline, from the moment you drop in a PDF to the moment the breakdown appears, with nothing hidden.
The pipeline, stage by stage
Worked example: a typical paper pasted or extracted to roughly 8,000 tokens of input (about 6,000 words, abstract plus body), producing a breakdown of about 2,500 tokens of output.
| Stage | Runs on | Cost | What happens |
|---|---|---|---|
| 1. PDF / Word upload | Your browser | $0 | The file is read and its text extracted on your own machine (pdf.js, mammoth). The file never leaves your browser, so there is no upload, no server storage, and no compute cost to anyone. |
| 2. Send text to the API | Vercel function | ~$0.00002 | Only the extracted text is sent. A serverless function runs for a few seconds. At Vercel's rates this is a fraction of a cent per request, usually inside the monthly free allotment. |
| 3. Read the paper (free tier) | Open-weights model | ~$0.007 | Llama 3.3 70B via Hugging Face Inference Providers. 8,000 input + 2,500 output tokens at roughly $0.60 / $0.80 per million. This is the only real cost, and it is under a cent. |
| 4. Render the result | Your browser | $0 | The breakdown is JSON, rendered client-side. Sharing, PDF export, and saving all happen on your device. |
Bottom line, free tier: a typical paper costs about 0.7 of a cent in model inference, and up to roughly 2 cents for a long one. Everything else rounds to zero. To you, the reader, it is free.
If you bring your own key
The optional deep read uses your own provider account, so it costs the site nothing and costs you whatever that provider charges. Same worked example (8,000 in, 2,500 out). Provider prices are set by the provider and change over time, so treat the non-Claude figures as approximate.
| Engine | Rate (in / out per 1M) | This run | Note |
|---|---|---|---|
| Free tier (Llama 3.3 70B) | $0.60 / $0.80 | ~$0.007 | Good, honest read. The default. |
| DeepSeek | ~$0.27 / ~$1.10 | ~$0.005 | Cheapest BYOK option. |
| Gemini 1.5 Pro | ~$1.25 / ~$5 | ~$0.02 | Has a free tier that often covers a few papers. |
| OpenAI GPT-4o | ~$2.50 / ~$10 | ~$0.045 | Approximate; set by OpenAI. |
| Claude Opus 4.8 | $5 / $25 | ~$0.10 | Sharpest deep read; highest cost per paper. |
Why so cheap
- The heavy file work (parsing a PDF) happens on your device, not a server.
- Nothing is stored. There is no database, no account system, no ongoing cost per user.
- A paper is small. Even a long one is a few thousand tokens, and modern inference is priced per million.
- The curated library was generated once, offline, and is served as static files.
All figures are estimates for a typical paper and depend on length and current provider pricing. The point is not the exact cent, it is that reading a paper this way is close to free, and now you can see exactly why.