Kaptr.me
Sign in Start free

Kaptr AI · Benchmarks

Kaptr AI,
in numbers.

Three AI models, each for a different job. Here is how long each one takes, how often it works, how good the answer is and what it costs, all measured on real runs.

Updated 7 October 2026Fast, Vision or Agent?What is Kaptr Agent?How Ask AI worksHow we measured

QualityAnswers rated good or excellent · higher is better
SpeedMedian seconds per answer · lower is better
PriceCredits per run · lower is cheaper
Quality vs. costQuality against credits per good answer (log scale). The dashed line is the value frontier: the best quality you can get at each price.
Cost, rankedCredits per good answer, cheapest first.
Every answer, gradedHow the AI check rated each run. Hover a segment for the count.

Fast, Vision or Agent?

Same goal, three ways to get there.

All three find the page that answers you and capture the exact part that matters. Fast and Vision even share the same framing rules. What changes is how many options they look at and how carefully the AI judges them.

Kaptr Fast

Speed first
Looks at
A few top sources, and stops about 1 second after the first good capture
AI judge
A compact model compares up to 4 captures
If nothing fits
One quick retry on other sources
Memory
Remembers the sites that answered well, and opens them first next time

8.7 s median · 18 of 20 good or excellent

Best for: quick checks on well-known pages, like a price, a score or a weather forecast.

Kaptr Vision

Most reliable framing
Looks at
Twice as many sources, and waits for at least 3 of them before choosing
AI judge
A larger model with a second opinion compares up to 8 captures, 2 per site
If nothing fits
A second search round on sites it hasn’t opened yet
Patience
More time per page to get past cookie banners and wait for charts to load

14.2 s median · 20 of 20 good or excellent

Best for: the exact table, chart, heatmap or ranking you want to keep on your board and refresh.

Kaptr Agent

Does the work for you
Looks at
Opens sites and uses them: searches, clicks, filters, scrolls
AI judge
A verifier checks every criterion in your request before accepting a capture
If nothing fits
Moves on to reserve sites
Refresh check
Reopens the page to make sure the capture can refresh

46.2 s median · 3 of 3 excellent

Best for: answers no single page shows: listings, offers or products that must match your criteria.

Each model is graded by its own AI check, and Vision’s uses the larger model. Vision’s edge is consistency: in our 20 tests, every Vision answer was good or excellent.

Kaptr Agent

It does the browsing for you.

Fast and Vision find a page that already shows your answer. Some answers don’t sit on a single page: you have to open several sites, apply filters and check each result. That is Kaptr Agent’s job.

  1. 1

    Plans

    Turns your request into criteria (price, place, type…) and picks the sites worth opening.

  2. 2

    Browses

    Opens several sites in parallel and navigates them like a person would: searching, clicking, scrolling.

  3. 3

    Checks

    A second AI checks the result against each of your criteria before accepting it.

  4. 4

    Captures

    Saves the result to your board, then reopens the page to make sure it can refresh.

A real run, second by second“Immobilier Dubaï 2 millions de dirhams” · 32.8 s in total
  • Price ≤ 2,000,000 AED
  • Located in Dubai
  • For sale, not for rent

Result: a capture of matching listings on squareyards.ae, all three criteria confirmed.

What you can ask it

  • Property“Apartments for sale in Dubai under 2M AED”
  • Standings & rankings“Full Ligue 1 table”, captured with all 18 teams in 46 s
  • Products & prices“Electric bikes under €1,500 with a 100 km range”
  • Listings with filters“Remote React developer jobs posted this week”
  • Offers to watch“Used Toyota Yaris under €12,000 near Lyon”

The first two were part of our benchmark runs. The others are examples of the same kind of task.

What it never does

  • Type a password, personal details or payment information
  • Buy, book, sign up, send a message or download a file
  • Leave public websites

Very heavy travel sites can take longer than the time Agent allows for a run. Kaptr then keeps what it already accepted.

Which one should I pick?

Start from what you need.

“I want a quick answer”Kaptr Fast~9 s · 5 credits “I want the exact table or chart, framed right every time”Kaptr Vision~14 s · 15 credits “I have criteria: price, place, dates…”Kaptr Agent~46 s · 50 credits

Ask AI · Beta

Ask your board a question.

Ask AI reads the screenshots on your board and answers in your language, citing the snapshot each fact comes from. It only answers from what your board shows. If the evidence isn’t there, it tells you instead of guessing.

  1. 1

    Open

    Click Ask AI at the top right of a board, or right-click a snapshot to ask about that one only.

  2. 2

    Choose

    Send every screenshot on the board, or pick only the ones your question needs.

  3. 3

    Ask

    Type your question, choose a model and, if you like, turn on Thinking to cross-check the values.

  4. 4

    Read the sources

    The answer lists its sources. Hover one to highlight the snapshot on your board.

Any anomalies worth checking?
Grounded answer

The conversion rate on S2 dropped to 1.8%, while traffic on S4 stayed flat. Worth checking the checkout page.

Sources S2 · Sales dashboardS4 · Analytics

Illustration of the answer format. The figures are made up for this example.

Three models, one fixed price per question

ModelCreditsThinkingScreenshotsPlans
  • Same price whatever the number of screenshots.
  • If the model fails, Kaptr tries another one at no extra cost.
  • No usable answer, no credits charged.
  • Business: use your own AI key, at 0 credits.

Screenshots are cropped, resized and compressed, then processed by a remote AI service. Don’t use Ask AI on content you wouldn’t send to an external AI.

How we measured

The fine print, in plain words.

Every number on this page comes from logged runs, in the data file below.

  • Speed is the median time from sending the question until the capture is checked and saved: half the runs were faster, half slower.
  • Price is the number of credits charged per run. Credits per good answer divides it by the share of good or excellent answers.
  • Quality comes from Kaptr’s automatic AI check, not a human review. Excellent: the capture shows exactly what was asked, tightly framed (for Agent: the verifier accepted it against your criteria). Good: it answers the question with a looser frame. Weak: a fallback capture of the page.
  • Runs: Kaptr Fast and Kaptr Vision were each asked the same 20 questions on 7 Oct 2026 (stock markets, weather, sport, crypto, games, prices), one after the other, on our production cloud browsers and AI models. Kaptr Agent: 3 runs on 2 tasks (30 Sep 2026); more Agent runs are coming.
  • Versions measured: the versions available today: Kaptr Fast 2.1, Kaptr Vision 2.2 and Kaptr Agent 1.
See every run
ModelQuestionTimeResult

Download the data (JSON)

Try it on your question

Ask once. Keep it live.

Pick Fast, Vision or Agent, type your question, and Kaptr puts the answer on your board, ready to refresh.

Start freeSee plans