How accurate is Ania?

The score

How fast is Ania?

The monthly re-certification runs the real load test — a ramp to 150 users in both write modes, plus the certified sudden-surge shape — and publishes the 95th-percentile latency here, month by month. Lower is better; the monthly gate is 2s. Surge timing is the time to first reply under a 200-user spike.

Every message, every answer

Open any row to see the exact test message and Ania's actual reply. Nothing is hand-picked — this is the whole table, in the fixed order the benchmark defines.

How Ania checks

Step 1 — the rulebook (every plan). Every message Ania reads goes past a built-in rulebook: years of known scam patterns — fake bank alerts, parcel-fee demands, lookalike web addresses, "act now" pressure — checked in an instant. It runs for every family on every plan, and Ania's promises never change with your plan: she will never ask for your password or a verification code, and she will never tell you to send money.

Step 2 — a second reader (paid plans). When a message is genuinely tricky — the rulebook isn't sure, or you send a screenshot to read — paid plans add a second pair of eyes: a powerful AI model that looks at the whole picture and can catch brand-new tricks no rulebook has seen yet. One rule stands above it: the second reader can only make Ania more careful, never less. If the rulebook says "scam," nothing talks it down.

What that means for you: the free plan is the full rulebook on its own. Paid plans are the same rulebook plus judgment on the hard cases — never a different set of rules.

How to read this honestly

What this measures: the Ania Gauntlet — 147 fixed messages across these categories: obvious scams, ambiguous messages, claims that point somewhere Ania couldn't verify, legitimate messages, tech problems, and break-it attempts (people trying to make Ania break its own rules). Each has an expected outcome defined in advance, and the grader is the same code our automated tests run on every change. The test messages are in the public repo — you can rerun the whole benchmark yourself.

What this doesn't measure: the endless variety of real scams. These 147 are a floor, not a ceiling — a fixed regression set that catches when an update makes Ania worse. New scam patterns appear every week; that's exactly why Ania combines the rule engine with a live AI layer and why the honest answer to "is this safe?" is sometimes "here's how to check with the source that can't be faked."

The numbers shown are from the rulebook alone — the offline rules engine from step 1 above, with no AI in the loop — so this page measures exactly what the free plan gives you. How Ania works →