Cloud LLM Setup: Which Model Fits Your Data, Law and Volume
The cloud model under your bots and agents, chosen on your own questions, set up in your company's name and kept switchable.
If your company builds or runs its own AI, the cloud model under it is a choice you will live with: it sets the quality of the answers, what happens to your data and the monthly bill. We choose it on your own questions rather than on a leaderboard, set it up in your company's name and prepare the way out, so moving to a better or cheaper model later starts with re-running a test, not with a rebuild.
Whether you need the cloud at all or a closed network on your own servers is answered on the AI platform page. This page is for when the answer is the cloud: which way to get a model, which model, what it costs to run and what a switch takes.
Three ways to get a cloud model
| Do it yourself (an API key) | Cloud setup by Ksentra | A chatbot or agent from us, on our infrastructure | |
|---|---|---|---|
| Model choice | your developer picks and tests it | 2–3 candidates tested on 20–30 of your questions; the test set stays with you | ours, tested on your questions |
| Knowledge base and search quality | you build and test it | built, and search quality checked on your test set before go-live | built in |
| Account and data | yours | in your company's name; request logging switched off where the vendor allows it | on our infrastructure |
| Switching vendor later | rework | prepared at setup: a test set to re-run, the indexing model chosen separately | our job |
| Cost control | the vendor's bill | a monthly cost-and-quality note | in our fee |
| You pay | tokens | setup and support to us, tokens to the vendor | the chatbot or agent price |
| Fits | one simple experiment with your own developer | several AI projects, your own developers, data kept in your accounts | one chatbot or agent, no IT team |
We have built this layer inside our own chatbot engine: request logging is switched off on every model call it makes, and its token use is measured below.
Cloud models compared
A token is a short piece of text, usually part of a word. Vendors charge per million tokens: for the tokens you send (input) and the tokens the model writes (output). List prices as of 21 October 2026:
| Class | Example model (vendor) | Input / output per 1M tokens | Where requests are processed |
|---|---|---|---|
| Most capable | Claude Opus 5.5 (Anthropic) | $4 / $20 | the vendor's cloud; regions per the vendor's terms |
| Balanced | Claude Sonnet 5.5 (Anthropic) | $2 / $10 | the vendor's cloud |
| Balanced | Gemini 3.8 Flash (Google) | $0.75 / $3.75 until 31 December 2026, then $1.50 / $7.50 | the vendor's cloud |
| Low-cost | Claude Haiku 5.5 (Anthropic) | $0.10 / $0.50, prompts up to 100,000 tokens | the vendor's cloud |
| Low-cost | Gemini 3.5 Flash-Lite (Google) | $0.30 / $2.50 | the vendor's cloud |
| Open model you host | for example Qwen3.6-27B (open weights, Apache 2.0 licence) | no per-token price: one rented 80 GB GPU server, about $1,290 a month, paid to the host directly. This is the closed-network route: setup from $10,000 (see the AI platform page) | the server you rent, in the region you choose; you can move it to another host |
The table is a sample, not the whole market: OpenAI and other vendors are compared the same way. Prices change, so we re-check them before every recommendation. Vendors' region terms are a requirement for us, and we pick models that meet them.
A leaderboard picks the candidates, and your questions pick the model. An LLM leaderboard ranks models on someone else's tests. In a setup we take two or three candidates and run them on 20–30 questions from your own work, each answer checked against your knowledge base. The winner is the model that answers your questions well at a price your volume can carry.
What a company chatbot costs to run
The model's bill depends on five things: chats a month, the size of your knowledge base, the model class, the number of channels, and whether a second model stands by.
Running cost a month ≈ chats × tokens per chat × price per token + knowledge-base hosting.
On the chatbots our engine runs (mostly scripted dialogues, with a model for free-text questions), a chat used about 2,200 tokens at the median and about 8,000 or more in the busiest tenth; about 94% of them were input (83 chats, mostly in Russian, July–October 2026). Token counts differ between models, so read these as the order of magnitude; the pilot measures yours. A chatbot that sends every turn to the model uses more.
As an example at list prices, 5,000 chats a month come to about $1.40–5 at a low-cost model's price (Claude Haiku 5.5), $10–37 at Gemini 3.8 Flash's (about $20–74 from January 2027) and $27–98 at a balanced model's (Claude Sonnet 5.5).
| Minimal configuration, 5,000 chats a month (estimate, as of 21 October 2026) | Month 1 | A steady month |
|---|---|---|
| You pay Ksentra | setup from $3,000 + support from $500 | support from $500 |
| You pay the model vendor | $1.40–98 a month in tokens | $1.40–98 a month in tokens |
| You pay the cloud provider | knowledge-base hosting and cloud resources, sized in the spec | the same |
| Total for 5,000 chats, before hosting | from about $3,500 in month 1 | from about $500 a month after |
Two configurations to choose from:
- Minimal, enough for a pilot: one channel, one model, no standby model, a monthly check. Request logging is switched off where the vendor allows it, in both configurations.
- Optimal, for production: all your channels, a second model on standby, continuous monitoring and a monthly report. It needs support above the floor.
Support from $500 a month is the floor: a monthly check, the monthly note and occasional knowledge-base updates, with replies within business hours and no uptime guarantee. Continuous monitoring and model switches are quoted above it or billed hourly.
Agents that read many documents per task use far more tokens; the sizing spec runs the numbers for yours. The spec costs $1,500, counts toward the setup fee and is optional on the cloud path.
Switching models later
- The connection. Many open-model hosts and several vendors accept the same request format. Where the new model does, the connection changes in configuration. Where it doesn't, or your bot relies on tool calls or structured answers, we adapt the connector. Either way we re-run your test set on the new model before it goes live, and you approve the result.
- Prompts. We re-tune and re-test them, because answers change between models even at the same quality.
- The model that indexes your knowledge base. Changing it means re-indexing the whole knowledge base. This is the larger job, so we choose that model separately and change it rarely.
Customer personal data and a foreign model
A privacy gateway is a server in your jurisdiction between your bot and a foreign model. It finds personal details in each request (the user's message and the knowledge-base passages sent with it), replaces them with placeholders, sends the cleaned text out and puts the real details back into the answer. The table that links placeholders to people stays with you.
Detection can miss some details: rare names, an address written inside a sentence, unfamiliar number formats. It does not catch details that identify a person only in combination, such as a job title, a town and a date. The model can also rewrite or drop a placeholder, and it cannot reason about what a placeholder hides, so data the model must reason about stays on a model in your jurisdiction.
Frequently Asked Questions
Which model fits us?
The one that answers your own questions well, within your data rules, at a price your volume can carry. We test two or three candidates on 20–30 of your real questions and write the result down with one or two alternatives and their monthly costs. You get it in the sizing spec or as the first step of the setup.
What does a switch to another model involve?
A connection change (configuration where the request format matches, an adapted connector where it does not), re-tuned prompts and a re-run of your test set. If the model that indexes your knowledge base changes too, the knowledge base is re-indexed, which is the larger job. We tell you which case you are in before you approve the switch.
Can our customers' personal data go to a cloud model?
A cloud model hosted in your jurisdiction keeps it there. For a foreign model, the vendor's region terms are a requirement we plan around, and the privacy gateway reduces what the model sees. Where we process your customers' data ourselves, we agree a data processing agreement with you before that processing starts.
Why not just buy an API key?
For one simple experiment with your own developer, an API key is enough. Once you run several AI projects, need to show where the data goes or want to control the monthly cost, the setup pays for the model choice, the knowledge base, the tests and the way out.
Will model prices stay where they are in the table?
Not always, and some rises are announced ahead: Gemini 3.8 Flash doubles its list price on 1 January 2027. The monthly note shows what a rise does to your bill at your volume.
Price
Cloud setup
from $3,000
setup · support from $500 / mo
- 2–3 models tested on your questions
- Knowledge-base infrastructure
- Cloud account in your company's name
- Connected to your chatbots, assistants, agents and systems
- Pilot of about 2 weeks on your own documents
You pay Ksentra for the setup and support. You pay the model vendor and the cloud provider directly for tokens and cloud resources; we don't resell licences or cloud capacity. Half of the setup is paid when we sign, half at go-live. This is the model layer: chatbots and agents themselves are separate services.
Not sure which cloud model fits your data?
Tell us what you run and what your data allows, and we will show which models to test and what each costs a month.