24/7
Always on.
Nights, weekends, December. There is no meter to watch and no reason to switch it off at six. A slot is reserved for you, so a busy hour for someone else is not a slow hour for you.
Early access open
Your AI bill should
not be a surprise.
Most AI is billed by the word, so the more your team uses it, the less you can predict what it costs.
Why the bill never stays still.
Most AI services charge by usage. It sounds fair, until you see how usage really grows.
-
You pay for every word, going in and coming out.
AI services count text in small pieces called tokens. Every question you ask and every answer you get is measured and billed.
-
Each question re-reads the whole conversation.
The assistant does not remember on its own. Every time you write, the whole history is sent again, and billed again. The longer the conversation, the dearer the next question.
-
One task can mean hundreds of requests.
An assistant that works on its own reads, checks, retries and tries again, and each step is billed. One stuck loop at three in the morning becomes a line on the invoice.
-
So the tool gets rationed.
When every question costs money, people ask less, finance asks questions, and the AI you set up sits unused.
A budget needs a number you can know in advance.
A meter cannot give you one.
See your own bill.
Then see it stop moving.
Pick a situation or set your own numbers. We work out what usage billing adds up to at list prices, and how many slots you would need instead.
Usage billing
€0
a month, at list prices
NEURA Brain, 1 slot
Fixed
one fee, the same every month
Per-token prices are the OpenRouter list for Qwen3.8-27B on 11 October 2026: typical provider $0.15 per million tokens in, $1.875 out, $0.0375 cached; cheapest listed $0.04, $1.35, $0.018. Converted at €0.92 per $1. With caching on, 80% of the context is billed at the cached price. Slots are estimated at 15 seconds per request, a peak three times the average, and a slot busy at most 60% of the time. Your numbers will differ: this is arithmetic, not a quote.
What the price buys besides the arithmetic.
- A contract, not a terms page. Hosted in the EU, a DPA, no training. Most per-token routes do not tell you where a request lands.
- Capacity that is yours. A slot is reserved for you. No rate-limit surprises when the shared pool is busy.
- The same price when an agent loops. A retry storm at 3am is work, not an incident on the invoice.
- Memory that does not cut. The whole history is kept and brought back, not trimmed.
How many slots do you need?
A slot is one answer at a time. It is not a person, and it is not a seat.
- Ten people is not ten questions at once. They read, they think, they type the next one. For most of the minute nobody is waiting on the model. When one answer lands, the next colleague can ask.
- The busy second is about one in three. In a professional office the peak we plan for is 25-30% of the room sending at the same moment. The rest are still working, just not hitting the endpoint together.
- A slot is not a wall. A slot answers one request at a time. Send three together and they queue: the second starts when the first lands. At about ten seconds each, three requests together are all back within half a minute. If the wait shows up in the room, you add a slot.
- The number follows from that. 10 people × 30% = 3 questions in flight. Round up for a bad Monday: 4 slots.
10 people×30% peak=3 in flight→4 slots
Typical chat-assistant seats are priced per person, and they are not private by contract. A slot is shared by the whole desk.
Agents are the exception: they can stay on a question for hours. If the desk is agents, plan one slot per agent that stays hot, or one for every two that take turns.
So we removed the meter.
NEURA Brain is private AI for a fixed monthly fee. You reserve slots, a slot being one answer at a time, and use it as much as you like.
Fixed
A fee, not a meter.
Every word, every retry and every loop is included. Finance writes one line in January, and it is still true in December.
Private
Your data stays yours.
What you send answers you and builds your own memory, nothing else. Hosted in the European Union, operated by IANUSTEC S.R.L. in Italy, with a DPA.
Uncut
A long conversation costs no more.
The model reads long files in full. When a conversation outgrows it, IANUSTEC keeps the whole history and brings back what each answer needs, instead of cutting the oldest turns.
Tasks we run every day.
On our own systems.
These come from IANUSTEC's own assistant: Open WebUI with tools connected to Odoo, Nextcloud, email, calendar and a browser. Many tool calls, long results, private data. That is the work the endpoint is built for.
Mileage claims, in one request.
The assistant lists the month's calendar events, finds the company vehicle in Odoo, works out the trips and writes the odometer entries. One real conversation: about 90 tool calls, each result longer than the question that started it.
Real task, run in our assistantFind the message, the attachment, the folder.
It reads the inbox, opens a forwarded email and its attachment, and searches the shared folders for the matching practice. Documents that would never be pasted into a public chat.
Real task, run in our assistantEvery lead, one by one.
It reads the whole pipeline, counts by stage and revenue, and flags what has no activity. Measured in September 2026: 167 leads analysed in about two minutes, with correct totals in 9 runs out of 9.
Measured, September 2026Which tickets are late, and whose.
Overdue tickets, who has the most, the five most urgent with priority and due date. In our tests the dates came out right in 3 runs out of 3, on a list of 120 tickets.
Measured, September 2026An agent that works a web page.
It navigates, reads the elements on the page, clicks and fills in forms, one step after another. A single browser test in our assistant took 136 tool calls: exactly the kind of loop that is expensive on a meter.
Real task, run in our assistantAn assistant that remembers the practice.
One private conversation per client or case, growing for months: files, emails, decisions. The model reads the file directly. When the history outgrows it, IANUSTEC keeps the whole conversation and brings back what each answer needs.
Law firms, accountants, consultancies, local government
Measured in our tests
Numbers from the real stack.
Numbers from the real stack.
October 2026.
We test the service with real, long chats and reasoning switched on, before launch. Here are the figures, and the limits that come with them.
≈ 1 s
The model starts working.
On a conversation that is already open, with a long history behind the question, thinking begins in about a second (median 0.9 s, 95% of requests under 1.7 s). The written answer follows once the reasoning is done: typically 5 to 8 seconds.
60-120
Tokens a second, faster than anyone reads.
Between 60 and 120 tokens a second, about 45 to 90 words, depending on how busy the pool is. The answer scrolls as fast as you can follow it.
14 s
The first message of a long conversation.
Reading a 200-page conversation from scratch takes about 14 seconds. After that it stays warm, and every further message starts in about a second.
10,320
Answers under load. Recall 100%.
Earlier test, August 2026: 10,320 requests against a 166,000-token case file, every detail recalled, zero errors. On a paired statistical test (HotpotQA) the result was no different from giving the model the whole file.
| What happens | What to expect |
|---|---|
| A conversation you already have open | Thinking starts in about a second. Written answer in 5 to 8 seconds with reasoning on. |
| The first message after a pause or a restart | The conversation is read again: about 14 seconds for 200 pages. |
| More requests at the same moment than you have slots | They wait in a queue and start on their own. Past a maximum wait you get a clear retry response, never a silent failure. |
| Agents running back to back, without pauses | Tested: they keep writing at about 20 tokens a second or better. |
| The full service, with the IANUSTEC memory layer on top | Not in these figures yet. We re-run them on the complete service before launch. |
Serving engine measured 2-10 October 2026. Raw data under NDA, or we re-run it live on your own file, on a call. The method stays closed. The outcome does not.
Answered in Europe.
Never used for training.
European rules are not a page in the footer. They are how this is built. The work stays in the European Union. What you send answers your question and builds your own memory, and nothing else: no training, no dataset, no resale.
In the Union.
The answer is made in the EU. The country can change from one member state to another. The Union does not.
AI Act, by design.
IANUSTEC runs this, from Italy, and a named company is accountable. That is the starting point of the Act, not a sticker added after launch.
Memory that is yours.
Your conversation memory is kept for your account only, never used to train a model, and deleted on request or when the contract ends.
Where the work happens
Green is a country in the Union.
Every pin is in the Union. The work can move from one EU country to another. It does not leave. The United Kingdom is on the map, and it is not a pin: we do not leave the Union to save money.
Four ISO certificates.
So you do not have to worry.
IANUSTEC holds them, so security and privacy stay on our side of the desk. NEURA Brain is a product of IANUSTEC S.R.L., Italy.
The certificates are the audit. The contract is the rest: a DPA is available, and your data is deleted on request.
For developers
One line changed.
One line changed.
Everything else identical.
Any client with a base URL field works. Paste the URL, a key and the model name, and keep the tools you already use.
from openai import OpenAI
client = OpenAI(
base_url="https://api.neurabrain.io/v1",
api_key="nr-sk-...",
)
resp = client.chat.completions.create(
model="neura-brain",
messages=[{"role": "user",
"content": "What did we decide about the Rossi contract in March?"}],
)
print(resp.choices[0].message.content) # same SDK, same callcurl https://api.neurabrain.io/v1/chat/completions \
-H "Authorization: Bearer nr-sk-..." \
-H "Content-Type: application/json" \
-d '{
"model": "neura-brain",
"messages": [{"role": "user", "content": "What was the penalty clause we agreed in March?"}]
}'import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.neurabrain.io/v1",
apiKey: "nr-sk-...",
});
const resp = await client.chat.completions.create({
model: "neura-brain",
messages: [{ role: "user", content: "Pick up where we left off" }],
});
Base URL
https://api.neurabrain.io/v1
API key
nr-sk-…
Model
neura-brain
Chat apps
Agents that stay on
Plan one slot for each agent that stays hot around the clock.
These clients are free software, yours to install. We do not ship a chat app: NEURA Brain is the private endpoint they talk to.
Reserve your slot.
Pay nothing until it is live.
Early access is open. Reserve the slots you need, at a price that stays locked for your first twelve months. A person at IANUSTEC writes back within 24 hours.
Not users. A slot is one answer at a time. In an office about one person in three is asking at any moment, so ten people usually need three or four slots. An agent that stays hot around the clock needs one of its own.
slot
164, or type more
Early access, 1 slot
Discounted for
your first year.
your first year.
A founding price, locked in writing for your first twelve months. Nothing to pay until your slots are live.
Afterwards we tell you the regular price at least 60 days ahead, and you can leave.
- No seats. Ten operators or a hundred. You do not pay per person.
- 1 slot. The only thing you buy: how many answers run at the same time. Extra requests wait their turn.
- No token meter. Input, output, retries and agent loops are not counted. A fair-use policy protects the pool from abuse.
- Reasoning on. Every answer thinks before it writes.
- Memory that does not cut. The whole history is kept and brought back when an answer needs it.
- Private. Hosted in the EU, a DPA, never used for training.
- A person, 24 hoursNot a queue. Someone reads this and writes back.
- Nothing to pay nowYou pay when your slots go live, not before. Cancel any time before then.
- Your file, liveWe re-run the measurements on your own data, on a call.
Reserved. Someone at IANUSTEC will write within 24 hours. Nothing to pay now.
This page is the endpoint.
The infrastructure is NEURA.
IANUSTEC also builds NEURA, a complete private AI infrastructure: software and hardware, on-premises or in our cloud, three redundant nodes. The applications are already there and there are no SaaS seats on top.
Software and hardware
One system: chat, files, automations, admin and the machines they run on. Not a chat bolted onto someone else's cloud.
On-prem or cloud
NEURA Rack in your office: data never leaves the building. NEURA Cloud if you want the same platform hosted by IANUSTEC. Same stack, you choose the perimeter.
Redundant. Always on.
Three identical nodes, no master. One fails, the others keep the company running. IANUSTEC watches the cluster, never the conversations.
No application subscriptions
Open WebUI, Nextcloud, n8n and the admin panel are included, integrated and maintained. Odoo, OpenProject and Chatwoot when you want them. No per-app seats.
Under the hood
What NEURA Brain
What NEURA Brain
runs on today.
Said plainly: today NEURA Brain is Qwen3.8-27B, an open-weight model from the Qwen team. It is the best balance of cost and quality we have found for everyday operations, which is exactly the job of this service. And it is only the beginning: we have just started, and the model behind the name will change as better ones arrive.
Qwen3.8-27B.
A 27-billion-parameter model that reasons before it answers, reads long files and calls tools. It is the same open model you could download yourself.
- Open weights from the Qwen team
- Reasoning on, built for long files and tool calls
The service around it.
A model is a file. A service is a price, a queue, a memory and a contract. That is the part you buy, and it stays the same when the model changes.
- Flat-price slots, with a queue instead of refusals
- Memory that keeps the whole history, uncut
- EU hosting, a DPA, no training, ISO-certified operator
- An OpenAI-compatible API your tools already speak
Tested before we ship it: on our agent test battery (CRM analysis, ticket triage, coding tasks, 18 episodes) the build we serve scored no worse than the full-precision original. When a better model replaces this one we say so first, and we run the same tests.
Certificate