Cost analysis
The organisation sees what each Octy™ answer costs and pays its own model provider at its own rates. Octy finds the relevant passages without a model call, and calls a model only when an answer needs an explanation, choosing the cheapest model that meets the profile’s quality bar.
- A question arrives from the profile's audience, on one of its channels.
- Controls: who may ask, from where, how often, and within the spend cap. A question refused here costs nothing.
- Retrieval searches the profile's knowledge only, within the caller's access scope.
- Not covered: if nothing is relevant enough, Octy says so, and no answering model is called.
- Retrieval answer (Coming): the passages themselves, quoted with their sources, and no model is called.
- Explained answer: a small model writes the answer from the passages, through your gateway. The router choosing the cheapest model that meets the quality bar for each answer is Coming.
- Evidence: every source is cited with the answer. A signed receipt is Coming.
- Measurement: counts, timings and the cost of each answer go to LightHouse. No content does.
Retrieval, at almost no cost
Section titled “Retrieval, at almost no cost”The search over the profile’s knowledge runs inside the deployment, on a keyword index and a meaning index built and searched there. Its cost is the compute the deployment already runs. See How retrieval works.
Two kinds of answer come from retrieval alone:
- Not covered. When no passage in the knowledge is sufficiently relevant, Octy states this before an answering model is called.
- Retrieval answers. The passages themselves, quoted with their sources, answer many questions directly, such as “where is the policy on X” or “what does section 4 say”. A retrieval answer calls no model, and its cost is close to zero. Every profile, including Octy on these docs, uses retrieval answers.
Explained answers through the router
Section titled “Explained answers through the router”When a question needs an explanation that draws several passages together, a model writes the answer from the passages, through the organisation’s gateway. Octy in a deployment without Octy profiles names its models in its configuration: a small one that decides how to handle the question, and one that writes the answer. Only models that the gateway’s policy allows and prices, in the deployment’s region, can be named.
For an Octy profile, every explained answer goes through the router. It keeps the models allowed where the profile runs, drops those whose answers fall short of the profile’s quality bar, tries the cheapest first, and escalates only when an answer misses the bar. See The router for every profile.
An explained answer is a short prompt of a few passages and a short reply, run on a small model. It typically costs a fraction of a cent, estimated. Sending every question, with whole documents as context, to a large model costs many times more, estimated.
Who pays for the models
Section titled “Who pays for the models”Profiles in a client deployment call models through the client’s gateway, on the client’s provider account: Amazon Bedrock, Azure AI Foundry, Google Vertex AI, the Claude API, or an open-weight model on the client’s GPUs. The client pays its provider directly, at its own rates. Atlantic does not resell model usage. Atlantic-hosted profiles run on models in Atlantic’s account.
Every profile has a monthly cap on its model spend, reserved before each model call. See Controls.
Cost analysis per profile
Section titled “Cost analysis per profile”Octy’s model calls in a deployment are recorded by the gateway with the deployment’s other usage, and shown in LightHouse™ under cost and routing. See What is measured.
A cost analysis for each Octy profile, on its Costs and questions tab and, for every profile together, under Octy profiles → Costs, for the last 7, 30 or 90 days.
| Figure | What it is |
|---|---|
| Answers | Answers given in the period, split into those from retrieval alone and those that used a model |
| Without a model | The share of answers made from retrieval alone |
| Model spend | What the profile’s model calls cost, against its monthly cap |
| Cost per answer | Model spend divided by answers |
| Model mix | Each model’s answers, cost and share |
| Router decisions | The model chosen first, escalations, and why answers missed the bar. See Reading the decisions |
| Latency | The median and 95th percentile time to answer |
| First answer after a wake | For a deployment that sleeps between questions, the median and 95th percentile time of the first answer after it wakes, and how many wakes there were. See Deployment profiles |
| Estimated saving | What the same answers would have cost had every one gone to a large baseline model, less what they cost. The baseline model is named beside the figure |
| Answers per day | The daily count, split by how each answer was made |
The figures come from the counts, tokens and costs each answer reports: from LightHouse for an Atlantic-hosted profile, and from the deployment’s operational telemetry for a profile in a client deployment, or from its imported bundles when it is air-gapped. The figures do not use question or answer text. The API is GET /api/v1/octy-profiles/costs for every profile and GET /api/v1/octy-profiles/{id}/costs for one.
The saving uses the same method as the other savings LightHouse shows. See How a saving is calculated.
Every figure says estimated
Section titled “Every figure says estimated”A cost per answer is measured from the provider’s own prices for the tokens each call used. A saving is a comparison with a counterfactual. It is an estimate, and LightHouse labels it as one wherever it is shown. See Why every figure says estimated.