Kriton — Stop guessing which AI to ask

Why people use it

Six reasons, in the order people tend to notice them.

The main thing

You stop having to choose

Different AIs are genuinely better at different things — one writes well, another reasons through code, another searches the live web. Kriton reads your question and picks. You never think about it again.

Nobody else does this

Ask several at once, then choose

A single model is confidently wrong in the same tone it is right. For a decision that matters, ask five the same question and read the answers next to each other. Where they agree, you can relax. Where they disagree is exactly where the real uncertainty is — and that is the bit worth your attention.

It gets better

It learns which one you trust

Every time you pick a favourite answer, Kriton remembers. Ask something similar later and the AI you preferred goes first.

Already paying elsewhere?

Already buy AI directly? Use that

If you have a developer API key from OpenAI, Anthropic or Google, connect it and Kriton uses it first, so those answers cost you nothing extra here. Worth being plain about: a ChatGPT Plus or Claude Pro subscription is not this. Those work only inside that company's own app and cannot be connected anywhere — an API key is a separate thing you buy separately.

Money

You pay for what you use

Not $20 a month whether you use it or not. The subscription pays for the app; the AI itself is bought separately and only spent when you ask something. We publish exactly what we add on top.

It knows you

It remembers between conversations

Keep separate spaces for work, a side project, personal things. Each remembers what matters without you re-explaining it every time you come back.

Three ways to ask

Most of the time you just type and Kriton handles it. When a question deserves more, you have three buttons.

Compare

we call it the Council

Five AIs answer the same question. You read them side by side and keep the one you like. Best for decisions you would normally ask a friend about.

About 3× the cost of one question

Combine

we call it Fusion

Several AIs answer, and their answers are merged into one. Best when you want the most complete answer and do not want to do the comparing yourself.

About 3–6× one question · takes ~10 seconds to start

Break down

we call it Dissect

A big, many-part question is split into pieces, each piece goes to the AI best suited to it, and the answers are stitched back together. Best for plans and projects.

Depends on how many pieces it finds

Routing: Each prompt is classified into a category — code, creative, analysis, research, general — and dispatched to a model pinned to that category. Which model sits in each slot changes as the field does — it is chosen for the category, not fixed to a brand, and the one that answered is named under every reply. You can override per conversation from the model selector; pinning a provider you hold a key for spends your key rather than credits.
Compare — the Council: A fan-out to 3–5 seats in parallel, streamed as they land, each tile showing its own state so a slow or failed seat never blocks the others. Picking a winner writes to a preference store that biases future routing for similar prompts. Seats are configurable; the default panel is five and costs about 3× a single turn, but the five most expensive pickable models cost roughly 15×, which is why the estimate is computed from your seats rather than shown as a fixed number.
Combine — Fusion: A router pseudo-model: several models answer and a synthesiser returns one response. The model that answers is never the one you named, by design. It streams, but every chunk arrives together — measured at ~10s to first token — so the interface says so rather than pretending otherwise.
Break down — Dissect: Decomposition into subtasks, per-subtask routing to a specialist, then a stitching pass. The full trace is streamed — you see the subtasks, which model took each, and the stitching step as it happens, rather than a spinner.
Memory: Messages and documents are embedded and retrieved by vector similarity, scoped per workspace, optionally across all of them. Long conversations are distilled rather than truncated. History is trimmed on a quantised boundary so the prompt prefix stays stable — a moving boundary silently destroys prompt caching on exactly the long conversations where it pays most.
Tools: Live web search, a sandboxed calculator that parses rather than evaluates, and your own HTTPS webhooks. Webhook URLs are validated against private, loopback, link-local and cloud-metadata ranges before any request is made.
Keys and billing: Bring your own provider keys and they are used first, encrypted at rest; credits are the automatic fallback when a key is missing or exhausted. Or run entirely on credits and hold no keys at all. Billing uses the cost the provider reports for the request that actually ran — including when a failover chain substitutes a model — never an estimate.
Everything else: Per-workspace personas and system prompts, prompt workflows for review-type tasks, speech in and out, artifact extraction, conversation forking, export and import, and an Android client sharing the same engine.

What you are actually buying

You are not buying messages. You are buying AI time, measured in credits. “A question” is just our way of saying how big a typical one is — some cost less, some cost more.

1 credit = 1 penny of AI use. That is the whole conversion. No points, no tokens to learn, no tiers.
You are charged by size, not by count. A short question with a short answer costs less than a long one. Nothing is deducted for asking.
You always see the estimate first. Before you press Compare, the button tells you roughly what it will cost — worked out from the AIs you actually chose, not a fixed number.
You are charged what it really cost. Billing uses the figure the AI provider reports once your answer is finished, never the estimate.
Credits never expire. Unused credit stays yours.

What you do	Roughly
One ordinary question	about 1.7 credits
Compare — five AIs at once	about 5 credits
Your free 50 credits	~30 questions
…or, if you only ever Compare	~9 rounds

Measured on 2026-08-29: 25 real questions, five in each of the five kinds Kriton routes to, through the models it actually uses. Costs ranged from 0.4 to 2.9 credits depending on how long an answer the question deserved — which is why the app shows you the estimate before you spend, rather than a promise.

Free to try properly. $5 a month to keep going.

There is one paid plan, and it is not a feature ladder. Everything works on the free account. What the membership buys is 250 credits a month, and the option to top up beyond them.

Free

$0 — no card

50 credits of real AI, once, when you sign up — around 30 full answers, or 9 Compare rounds
Every feature. Compare, Combine, Break down, memory, personas, voice, tools — none of it is held back
No daily limit. Spend the credits in one evening or over a month; they are the only limit. One thing that is capped: a single free-plan message tops out around 8,000 tokens, about 30,000 characters
Have a developer API key? Connect it and keep going for free — note this is not a ChatGPT Plus or Claude Pro subscription
Credits never expire

Membership$5 / month
250 credits every month, included. Around 140 full answers,
              more if your questions are simple ones. They roll over and never expire
Want more? Top up at cost plus 20%. Same rate for everyone, no
              volume ladder, published on this page
Supports the free tier and the work of keeping the app running
Same features as free — because there was nothing to hold back
Cancel whenever. Credits you already bought stay yours

Hold a developer API key from OpenAI, Anthropic or Google? Connect it and Kriton uses yours first, with your credits untouched as a backup. A ChatGPT Plus or Claude Pro subscription is a different product and cannot be used this way.

Try it on a question you actually care about.

That is what the free credits are for — enough to use it properly, not a demo.

Start free — no card needed