Hey PH 👋 I got tired of AI speed claims I couldn't verify, so I built a site that measures 12 models (Anthropic, OpenAI, Meta, DeepSeek, Google, Z-AI, NVIDIA) against 25 everyday tasks — write an email, summarize a paper, explain a block of code — and shows the real recorded time, tokens, cost, and accuracy for every pairing. Nothing is measured live when you visit. A scheduled job runs each task against each model on a fixed cadence and records the response as it streams in — visitors replay that recording at the pace it originally arrived. I did the cost math on live-per-visitor calls first: at any real traffic it's hundreds to thousands a month and scales with visitors, not with value delivered. Recorded costs the same at ten visitors or ten thousand. A few things that surprised me building it: An output token cap was silently truncating long-form tasks — understated generation time on 27% of my first real recording batch, no error, just quietly wrong numbers. Free open-weight models (Gemma, GLM, Nemotron) sit on the same 25 tasks as $100/M-token flagships. Free is genuinely slower on some tasks — now that's a number instead of a guess. "Which model should I use" needed cost, speed, and accuracy blended, not just whoever's fastest. The top pick for one category was cheapest and second-fastest, not fastest outright. Every number is a recorded run, a scored rating, or a labelled placeholder — methodology page states which, per number. Would love feedback, especially if something looks off. howfastai.com

How Fast Is AI

Real recorded speed, cost & accuracy for 12 AI models

Real recorded speed, cost & accuracy for 12 AI models