We put AI models through the same screenshots and quiz pages Peeksolve solves every day. Every answer graded, every millisecond counted. Then run it yourself.
Official runs on PeekBench v1. Click a column to sort. Models marked “In Peeksolve” are the ones you can pick in the extension.
| # | Model | PeekScore ↓ | Accuracy | First word | Auto-fill page | Cost / 100 questions | Best at |
|---|---|---|---|---|---|---|---|
| Loading… | |||||||
Each car arrives at the moment its model's first word arrived, in real time — the median over 16 exam screenshots.
Top left is where you want to be: right, and there in under a second. Point at a model to find it.
Share of tasks answered correctly in each subject, in percent.
Every task is a screenshot or a quiz page exactly as the Peeksolve extension sends it — no clean API prompts, no tricks.
An exam question on screen, answered from the picture alone. We time the first word and grade the answer.
Whole quiz pages filled at once: radio buttons, dropdowns and number fields, the way the Agent does it. We time each page.
Ask the models your own question, practise on exam pages with the extension, or check that every shortcut works.
Your question or screenshot, answered by DeepSeek, Gemini and Claude side by side.
Ask a question → FreeExam-style pages for every shortcut, graded the moment you finish.
Pick a set → 1 minutePress each key once and watch it arrive. Nothing is sent while you test.
Run the check →Sign in with your Peeksolve account, pick DeepSeek, Gemini or Claude, and watch all 77 tasks get graded live. A run takes about half a minute and costs a few cents of your Pro credit.
Start a run →Screenshots shrunk like the extension does, quiz pages captured by the real extension with its field list and key badges. Same prompts, same settings.
Multiple choice counts by the option picked, numbers by value (“10'816”, “37,5 %” and “1/4” all count). An answer left out is wrong.
Every run starts on Peeksolve's own server, through the same gateway. Times are what your answers would take, not lab numbers.
Anyone with Peeksolve Pro or a Day Pass. You sign in with your Peeksolve account; the run uses your own Pro credit, usually a few cents (DeepSeek well under one cent, Claude Sonnet more).
The three you can pick in Peeksolve: DeepSeek V4.1 Flash (the standard), Gemini 3 Flash and Claude Sonnet 5.5, each with or without “Think before answering”. The leaderboard also lists other models we test officially.
Speed changes a little from minute to minute with the providers' load, and models are not perfectly deterministic. Run it again and you'll see the spread.
No. Every run uses the same fixed tasks. Your results page shows the model and the scores, never your account.
Yes. Every run gets its own link with the full result, every answer included.
Yes. The practice sets and the shortcut check are free and work with any Peeksolve plan. Running the benchmark and asking the models side by side use Pro credit.