Plain answers to what people actually ask about how fast, and how expensive, AI models are. Where these numbers come from, and what they do and do not tell you.
Discover AI Speed (the front page) picks one task and one model and races it against a typical human time: the fast, one-glance version.
Compare AI Models goes deeper: run two models (one Anthropic, one OpenAI) side by side on the same task, with tokens and cost for both. Start on the front page; move to Compare AI Models for the head-to-head.
Among the models measured here, Haiku 4.5 (Anthropic) and GPT-5.6 Luna (OpenAI) are built for quick, everyday answers and measure fastest on simple tasks.
Fastest is not the same as best. Give a quick model a specialist-level task and a slower, larger model will handle it better. That is what the Simple / Intermediate / Complex tiers are for: matching the task to a model built for it, rather than always reaching for the biggest one.
Every provider charges per token, and the price per token differs a lot between models. A top-tier model like Fable 5 can cost ten times or more what a fast everyday model costs for the exact same task.
The token count shown is roughly the same across models for a given task; it is the price per token that changes the bill. That gap between "same task, same tokens, different price" is exactly what the Tokens & Cost numbers on every result are showing you.
No, and that is worth being plain about. A scheduled job asks each model each task once a week and records the answer as it streams in: every piece of text and the moment it arrived.
Pressing Race (or Run) replays that recording in your browser at the pace the words originally appeared. Nothing is generated while you watch, and your browser never contacts any AI provider directly for these measurements. The one exception on the site is the "Ask about this site" chat bubble, which does call a live model to answer your question, but it just never touches the speed, cost, or accuracy numbers themselves.
Because it is a published figure, not a fresh measurement: last week's recorded run, replayed now.
A run that looks wildly different from the previous week is reviewed before it is published, rather than shipped automatically. A provider incident or a network hiccup should not quietly become "the new normal" on the site.
Accuracy cannot be read off an API response; it has to be judged. Each output is checked against a fixed rubric written for that task, either by a human reviewer or an AI judge, and the card always says which method was used.
Where a task has not been scored yet, the site shows a clearly labelled placeholder rather than guessing.
No. Speed and cost say nothing about whether an answer is correct, useful, or well written on their own: that is what the separate accuracy score is for.
A fast, cheap, wrong answer is still wrong. Treat speed, cost and accuracy as three separate facts about a model, never collapse them into one score.
Every task uses a fixed prompt that is part of the site, so you do not type anything and there is nothing of yours to send.
Your browser makes no AI request at all; it replays recordings that already shipped with the site. The only thing that ever contacts an AI provider is the monthly recording job, and it runs on a schedule, nowhere near your browser.
Recorded monthly. The exact date a figure was measured travels with it: check the methodology page for the current run counts.
A run that fails a sanity check against the previous week is investigated rather than merged, so a bad measurement should never quietly reach the site.
Nobody pays for the numbers, and no sponsor influences what is measured or reported. If a sponsor ever appears, it will be labelled as one.
The site measures models from every provider (Anthropic, OpenAI, Meta, DeepSeek, Google, Z-AI, NVIDIA, MiniMax) side by side using the same tasks and the same method, so the comparison is apples-to-apples rather than home-field advantage for any of them.
It depends which model on each side, and on the task: there is no single answer that holds for every pairing. A small, quick model like Haiku 4.5 or GPT-5.6 Luna will usually beat a large specialist model from the other provider on an everyday task, and lose to it on a hard one.
Rather than one headline number, race the exact pairing you care about: pick a task on the front page, switch the model, and watch the two times side by side, or open Compare AI Models to run a Claude model against a GPT model directly, with the same task, same prompt.
There is no single per-message price: cost is tokens read and written, multiplied by that model's price per token, and both numbers vary. A short question costs a fraction of a cent on nearly any model; a long, reasoning-heavy answer from a top-tier model can cost meaningfully more.
What surprises most people is not the total, it is the spread: two models can use almost the same number of tokens for the same task and still land a bill ten times apart, because the price per token differs that much between providers and tiers. Every result on this site shows the actual token count and dollar cost for that exact task and model, not an average.
For the kind of short, well-defined tasks measured here (writing an email, summarizing a document, checking grammar), AI models typically finish in seconds what takes a person minutes, often 50 to a few hundred times faster depending on the task and model.
That gap narrows or reverses for specialist, multi-step work, which is the point of the Simple / Intermediate / Complex tiers: speed is not a fixed multiplier, it depends on how well the task matches the model. Pick any task on the front page to see the real recorded gap for that one, rather than a single "AI is Nx faster" headline.
← Discover AI Speed · Compare AI Models → · How we measure → · Model comparison → · Claude vs ChatGPT speed → · Paid vs open-source AI → · Cheapest for students → · Fastest for coding → · Voice AI → · Glossary →