Latest releases

15 SEP 2026

The safety check, redaction and document reading run on our own hardware in the EU, the platform speaks Danish, and subscriptions and payments are handled in the app.

SECURITY

The safety check runs on our own hardware

Every message and every answer is checked against our rules for use by a model we run ourselves, on our own hardware in the EU. Nothing leaves the platform to be checked, and Mistral no longer performs it. The check costs no tokens either.

A busy moment does not sign you out

When your session hits the limit on how often it may be renewed, the page tries again instead of sending you to login.

Signing out covers the whole browser

Sign out of chat and you are signed out of administration in the same browser too, and the other way round. Your other devices are not touched.

A slow server does not become an error page

When the server answers slowly, the page waits and tries again before reporting a failure.

CHAT

The platform speaks Danish

Choose Danish or English under Settings. Chat, settings, error messages, the sign-up email and every amount follow, written the way the language reads them.

Chat works on a phone

Dialogs can be closed, buttons are easy to hit, and if a message is lost on the way you send it again with one tap instead of typing it over.

Scanned pages and tables are read

A scanned page, a photo of a document or a table that cannot be lifted out as text is now read on our own hardware and passed on as text. A table reading is checked against the page's own text, so a number that is not on the page does not get through.

You can watch the document being read

Chat shows that attached files are being read instead of waiting without saying anything. Large documents get through faster than before.

Switch memory off for yourself

You can switch memory off for your own account without an administrator changing anything for the rest of the organisation.

REDACTION

Redaction runs on our own graphics card

The model that finds personal data now runs on our own graphics card, so a long document waits less. The result is the same: We compared the findings across both setups without finding a single difference.

What the tools are asked to do is redacted too

When the model passes on a search or another tool call, the content is redacted the same way the message itself is.

ADMIN

Change plan without leaving OurAI

Switch between Pro and Max straight from the account page. An upgrade applies at once, a downgrade from the next period, and you can undo it until it takes effect. Cancelling, and undoing a cancellation, live in the same place.

Buy tokens and replace your card in the app

Extra tokens are bought with the card in the app, and the card can be replaced in the same place. Every purchase gets an invoice, and the account page lists your payment history.

You are told when a renewal fails

When the card declines a renewal, the account page says so and offers to replace the card. The payment is completed on the new card at once, and your tokens follow.

Rename your API keys

An API key can be given a name, and the list shows its start and last four characters so you can tell them apart.

03 SEP 2026

Chat gets projects with pinned conversations, redaction catches API keys, and Gemini 3.8 Flash takes over.

CHAT

Projects own their conversations

The sidebar is rebuilt. Drag a conversation into a project, pin the important ones to the top, and rename with one click. Delete a project and its conversations go with it.

See what the model did

The timeline of searches, tools and reasoning is saved with the answer. Open an old conversation and it is still there.

Settings as a dialog

Settings open on top of the chat instead of sending you to a new page. Account deletion lives in the same place.

REDACTION

API keys and access tokens are hidden

Keys from known providers, JWT tokens and private keys are removed before the message reaches the AI model. On for everyone, whatever the plan.

Pause your own terms

Administrators can switch a custom term off without deleting it. The term stays in the list but is not redacted until it is switched back on.

MODELS

Gemini 3.8 Flash replaces 3.7

Same price, smarter model, and it scores higher on our Danish test. Gemini 3.7 Flash is switched off. On Basic, 3.8 Flash is the model you start with.

Claude Fable 5 is switched off

Anthropic keeps everything sent to that model for at least 30 days and runs it through their safety review across requests. Content that gets flagged can be held back for a safety investigation. That does not fit our data processing agreement. Claude Opus 5 and Sonnet 5 are not subject to that review and remain.

GPT-5.6 and Claude Sonnet 5 cost fewer tokens

The providers cut their prices, and that passes straight through to what an answer costs you.

Kimi K2.5 is switched off

Moonshot has retired the model. Kimi K3 and K2.6 remain.

01 SEP 2026

Use OurAI from your own code, and one page for data access and erasure.

CHAT

Use OurAI from your own code

Settings now mints API keys for calling the models from your own scripts and programs. A key follows your account, so redaction and usage work as they do in chat, and you can revoke it with one click.

Search reads more sources

Web search now reads PDFs and pages that could not be fetched before, and brings home far more text per search.

Errors say what happened

When something fails, the message now says what went wrong and what to do, not just that something went wrong.

ADMIN

One page for access and erasure

The Compliance page gathers data-access requests, erasure with a receipt, and the record of processing activities. Export one member's data or the whole workspace, and see what is covered.

The audit log knows every outcome

Paused and disconnected turns show as what they are instead of as failures.

24 AUG 2026

A new personal account gets the Pro features for 30 days, and Redaction now follows the plan in two rungs.

REDACTION

Two rungs instead of one

Pro swaps out personal details: names, emails, phone numbers, account numbers. Max and the company plans also take business details such as company names and CVR numbers, and the terms you add yourself. Change plan and your settings stay put, working again the moment the plan covers them.

Web addresses and access codes are always covered

They are technical identifiers rather than a business detail, so anyone with Redaction switched on gets them swapped out.

The shield in the composer answers whether Redaction is up

It reads both the plan and the setting your administrator chose, not just one of them.

CHAT

30 days of Pro features on a new personal account

Redaction, Auto model picking and image generation are included for 30 days from the moment you are set up. A counter shows how long is left, and you are told when it ends.

ADMIN

The policy page shows what the plan unlocks

Settings your plan does not cover are shown locked on the page rather than missing from it.

The out-of-tokens message names the person who can buy

Only the owner can buy tokens, so the owner is who the message now names.

Sign-in goes straight to the login page

No pages in between, and the whole wait is spent in the button.

22 AUG 2026

Auto picks the model, all content is encrypted, and four models join.

CHAT

Let OurAI pick the model

Choose Auto and OurAI reads the task in your message and picks the model to match. The pill above the answer names the model that replied, and tapping it shows why. Auto is Pro and up.

Answers stay

Ask for the answer again and the old one does not disappear. You can page between them.

The model can correct what it remembers

Instead of stacking duplicates, it can correct and delete what it has stored about you.

MODELS

Four new models, all measured in Danish

Every model in the catalogue is measured on Danish tasks, so Auto picks one that actually writes it well.

DeepSeek is switched off

They train on what is sent through their API, which our own data-handling commitments do not allow.

SECURITY

Conversations and files are encrypted in the database

Messages, what the model remembers about you, and the files you upload are stored encrypted. Each organisation has its own key, so a copy of the database cannot be read without it.

Links in answers are checked

Against a list of known malicious sites.

REDACTION

Attached documents are read with tables and equations intact

Which means fewer false alarms in Redaction.

17 AUG 2026

Guidelines on every message, your own terms in Redaction, and fewer sign-ins.

SECURITY

Guidelines on every message

Every message is held against OurAI’s acceptable-use guidelines before it reaches a model. What is judged is the intent, not the subject.

You are signed out less often

A gap of up to 14 days no longer costs a fresh sign-in.

REDACTION

Your own terms, always removed

Administrators can add up to 50 words and phrases that are always removed, under a label of your own.

A file that cannot be read by machine

A scanned contract or a photographed ID is not sent until you have seen it and said so.

ADMIN

Rules switch on and off row by row

Instead of one switch for the whole of Redaction.

09 AUG 2026

Web search reads the pages, Insights is rebuilt, and five questions pick the models.

CHAT

Web search reads the pages

It does not fetch headlines alone. Afterwards it tells you which pages it read, so you can see what the answer rests on.

Suggestions when a conversation is empty

And your message survives a reload while it waits for an approval.

REDACTION

Dates and ages come back in the reply

A date is hidden on the way out and put back in the reply, instead of being made imprecise for good.

ADMIN

Insights shows more

How much sensitive information is being removed, who in the organisation has got started, and when you will run out of tokens.

Five questions pick the models

Instead of knowing the models, you answer five questions about your work and get a list back.

30 JUL 2026

A small release.

CHAT

Edit an image instead of making a new one

You can ask for changes to an image you already had made.

A reference number when something fails

When an answer fails you get a number to send along when you write to us.

28 JUL 2026

Redaction catches more, you have the last word, and errors say what to do.

REDACTION

Redaction catches more

Organisation names, CVR numbers and clear-text passwords are removed by default, and dates and ages are made imprecise instead of disappearing: 28.03.2015 becomes marts 2015, so the reply still makes sense.

You have the last word

Keep everything, or keep one type of information, straight from the approval. The button always states what actually gets sent.

CHAT

Error messages say what you can do

Every error names what stopped the message and what the next step is. A refused first message no longer leaves an empty conversation in the list.

ADMIN

Governance gathers the rules in one place

Policies becomes Governance with a sectioned menu, and what your subscription does not cover yet is shown with a way there instead of hidden.

19 JUL 2026

The model remembers, five models join, and sign-in moves into OurAI.

CHAT

The AI model remembers what you tell it

Tell it something about yourself or your work and it is remembered and used in later conversations. What is remembered passes through Redaction every time, like everything else you write.

MODELS

Claude Fable 5 and Sonnet 5, GPT-5.6, Grok 4.5 and Kimi K3

Older versions overtaken by their own successor are switched off.

SECURITY

Sign-in, two-factor and password recovery inside OurAI

You are no longer sent to another site to sort out your access.

ADMIN

Repeated content is billed cheaper

Send the same long conversation again and you pay the provider’s lower price for the part they have already seen.

14 JUL 2026

Generated files are encrypted, the model choice sticks, and companies share one pool.

CHAT

A model choice that sticks

Switching model mid-conversation sticks for the rest of it, including when you come back from another device.

SECURITY

Generated files are encrypted

Documents and images made along the way are encrypted while they are stored.

ADMIN

Audit shows every round

The log shows exactly what was sent to the model in each round of a conversation.

One pool of tokens for the whole organisation

Tokens are bought separately and never expire.

04 JUL 2026

Documents and images, an account page for individuals, and switches for the administrator.

CHAT

Documents and images

You can have a document written or an image generated. It appears beside the conversation with versions and download. Image generation is Pro and up.

See what is happening while you wait

When checking a large file takes time, you can follow it instead of watching an empty screen.

ADMIN

Your own account page

Individuals get subscription, payment, profile and their own rules in one place in the app.

Switch individual features off

Administrators can switch off tools, including the ones that send data outside Europe, and decide which model new conversations start with.

26 JUN 2026

Redaction gets an approval step, and chat gets web search.

REDACTION

Approval before anything is sent

You see what was found, correct it, add your own terms and approve, before a single word reaches a model.

CPR numbers and seven other countries’ national IDs

Names are found by a language model rather than fixed patterns alone.

CHAT

Web search

The model can look things up when the answer needs something more recent than it knows.

Watch the model think

On the models that support it, you can follow the reasoning while the answer takes shape.

12 JUN 2026

OurAI goes live.

CHAT

Chat on every major model

The models from the leading providers in one conversation, on one subscription.

REDACTION

Redaction on every message

Sensitive information is found and removed before the message reaches a model.

ADMIN

Your own rules, Audit and Insights

Administrators decide which models may be used. Audit shows what was sent, Insights what it is being used for.