AI model controls

Models on Kariant carry Klick names — that is the name you pick in the model selector — with the branded engine behind each shown alongside it. The free defaults cover everyday work: Klick Fast and Klick Lite (Gemini Flash-Lite — Fast is the platform default), Klick Default and Klick Flash (Gemini 3.5 Flash), Klick Balanced (Gemini 3 Flash, the balanced thinking tier), and Klick Smart (Gemini Pro, with a 2M-token context window).

Premium models are metered in Koins and picked per run: Klick Depth (Claude Sonnet — deep strategic analysis), Klick Genius (Claude Opus — maximum reasoning depth), Klick Skorer (Claude Haiku — fast scoring), Klick Kode (Codestral — code generation), Klick Klassifier (Mistral Nemo — fast classification), Klick RAG (Command R+ — retrieval-optimized), Klick LongDok (Jamba — 256K-context documents), and Klick Validator (Llama — independent validation), plus Klick Video (Veo) and Klick Musik (Lyria) for creative generation.

Each entry is labeled with its provider, context window, and cost tier, so the choice is informed: a fast free default for routine copy, a premium deep-reasoning model for strategy work — per run, not per account.

Harder tasks deserve explicit reasoning depth. The thinking control offers none / low / medium / high levels — or a token-level budget you set yourself — and the extra thinking is metered transparently, estimated before you run. Thinking on free default models usually costs nothing extra; on premium depth models expect roughly 0.1–2.5 Koins (≈$0.10–$2.50) per 1,000 thinking tokens, with Klick Genius at the top of that range.

Context windows range from 128K tokens up to 2M tokens on the largest models. That means entire documents, catalogs, or account histories fit into a single run — no pre-summarizing. Context caching reuses long, repeated context across runs to cut both latency and cost on document-heavy workflows.

Live grounding can be toggled per run so answers cite current web sources instead of guessing. A grounded query typically costs ≈1.2 Koins (≈$1.20) on standard features, rising to ≈3.5 Koins (≈$3.50) on the most research-intensive features — always shown up front before you run. Generation parameters — temperature, top-p, top-k, and max output length — tune determinism versus creativity per task, and the response language is settable per run.

Two-stage Fusion combines both worlds: a grounding pass gathers fresh facts, then a depth model reasons over them — current information and deep reasoning in one run. It is on by default where it helps.

Every configuration shows its estimated cost against your live wallet balance before you run. Free default models stay free; premium models show a clear metered price — from fractions of a Koin up to ≈2.5 Koins (≈$2.50) per 1,000 output tokens on the deepest model — and thinking budgets and grounding show their meter up front. Kurator research features are priced at 1.5× the standard meter, and the estimate includes it. No configuration can surprise your wallet.

Agent and AI actions with flat rates are equally explicit: AI lead scoring is 1 Koin ($1) per scored lead (skipped honestly when the balance is empty — never silent free compute), a drafted sales-agent outreach sequence is 3 Koins ($3), and a browser-agent task run is 5 Koins ($5). At the agentic tier, your trained agents inherit the same controls — model, thinking, grounding — plus the training, evals, and guardrails you define, so autonomy never means loss of control or unbounded spend.

All charges are denominated in Koins (1 Koin = $1). Partners in India see the same balances and charges as Sikka, the ₹-pegged display currency — wherever a figure appears in headers, filters, or estimates, the India setting shows Sikka automatically.