Shrink the AI bill. People included.

VARIA is a decision-flow simulator for AI pipelines. You compose a flow of AI decisions: decision-model questions (Noul, Choice, Score), LLM steps, code rules, confidence gates and human handoffs. VARIA runs 10,000 items through it and shows cost per 1M items, p50 and p95 wait, decision calls per item and the share of work that reaches a person, before you build the pipeline.

Two groups first: engineers whose LLM classifies or routes and whose bill keeps growing, and heads of support or operations deciding how much of a queue to hand to AI. Agencies that sell automation use it to show a client the economics of a flow.

In the VARIA demo support flow (modelled), the machines cost about $0.0010 per ticket: one decision-model call plus an LLM reply for refund requests. People still touch 28% of tickets, 0.30 touches per ticket on average, so at $3 a touch the people cost about $0.90 per ticket. Wrong routes add about $0.42 if confidence is calibrated, for $1.32 all in, against $4.17 when people triage every ticket. Use your own prices; Gartner put a live-channel contact at $8.01 in 2019.

It depends on how your flow routes uncertain and sensitive tickets. In the VARIA demo flow, 72% of tickets finish without a person and 28% reach one: unclear team or low confidence (13%), urgent pages to on-call (14%) and refund replies held for review (3.0%). VARIA shows this split for your own flow before you automate.

List the steps, the tokens each step reads and the share of items that reaches each step. Cost per item is the sum of tokens × price × share, plus the per-call price of LLM steps times their share. Wait time comes from the mix of paths: add up step latencies along each path, then read p50, p95 and the mean across items. VARIA does this by simulating 10,000 items through the flow.

Jev by TypeSafe is priced at $0.042 per 1M input tokens; output is free. A support ticket with three questions in one request reads about 580 tokens, so a request costs about $0.0000244 and a million of them about $24.36. Several questions about the same text in one request are billed once for the text. VARIA is not affiliated with TypeSafe.

It multiplies the tokens each step reads by the public price ($0.042 per 1M input tokens for Jev; output is free), multiplies by the share of items that reach the step, and adds the per-call price you enter for LLM steps. Code steps cost $0. In the demo support flow that comes to $1,037 per million tickets: $27.49 of decision-model calls and $1009.20 of LLM replies.

No. Costs follow published pricing, but answer shares, confidence and latency are assumptions you set. The next step is calibration: run the flow on about 200 of your labelled items with your own key and replace the assumptions with measured shares and accuracy.

Observability measures a pipeline that already runs: traces, cost, latency and evals from real traffic. VARIA estimates a pipeline before it exists and lets you compare designs. They work together: model the flow in VARIA, ship it, then compare the expected shares with what your traces show.

A vendor’s calculator uses the vendor’s assumptions about their own product. VARIA models your flow with your prices and keeps every human outcome visible, so you see the work that stays with people next to the machine cost.

They are Jev’s three question types. Noul returns the probability of yes. Choice picks one of up to 255 options and returns a confidence. Score returns a level on a scale of 2 to 10 levels, with a confidence. Several questions about the same text can go in one request; they are answered in parallel and the text is billed once.

No. VARIA is an independent project. It models flows built on decision models such as Jev by TypeSafe and uses their public pricing. Jev is a trademark of its owner.

Yes. VARIA has an MCP server (tools such as get_flow, set_expected_share and add_step) and a CLI with a simulate command, so an agent can change the flow from a plain-English request and re-run the simulation. This website also exposes demo tools to browser agents through WebMCP.

A median moves only when half of the items change path. If a fast path takes 0.3 s and a slow LLM path 6.6 s, p50 stays at 0.3 s until 50% of items take the slow path, then jumps to 6.6 s. Always read p50 with p95, the mean and the share on the slow path.

What you send in the request form comes to the VARIA team and is used only to model your flow. The bench itself is a single HTML file that runs locally. Calibration, when it ships, will call the decision model with your own API key, so your examples go only to the provider you already use.

VARIA is in preview and has no public pricing yet. Send us your flow through the form: we model the first flows for free and send you the numbers. Access to the bench for design partners is planned.