Which models will be tested?

12 models from 8 providers are planned: GPT-6 Astra, Claude Opus 5.5, Gemini 3.1 Pro, Grok 4.7, Mistral Medium 3.5, DeepSeek V4 Pro, Qwen 3.8 Max, Muse Spark 1.3 as top models, and GPT-6 Luna, Claude Sonnet 5, Gemini 3.8 Flash, Mistral Large 3 as budget models. The list is not fixed; reference candidates are configured separately. No approved results have been published yet. Changes to the list are recorded in the changelog.

How often is the barometer updated?

The target is every week. No approved results have been published yet. Results are published only once a run has passed the automatic checks and Aprex has reviewed the outliers.

How are answers scored?

Each task scores 100 or 0. The answer must be JSON with exactly the required fields. Every field must match after normalisation (dates, amounts, digits, categories, booleans). Null is valid only where the key explicitly allows it. There is no partial credit in the v0.5 ranking; share of correct fields is diagnostics only. Missing answers are listed separately without 0. No AI judge.

Could the models have seen the tasks before?

The 12 example tasks on this page are public and may be known to a model. Calibration and holdout are not public: they are not shown on the site and are not in the website source or metadata. The repo is private without immunity to training, and a runner sends its assigned task text to the provider on each run. Holdout is reserved final evaluation. Results from public tasks and calibration are never combined into a claimed measurement on unseen tasks.

Why show cost and response time?

Because the best model is not always the best choice. For many tasks a cheaper, faster model may be good enough. Cost is the measured API price from the provider and is never mixed into the score. The subscription pilot reports no API cost. API and subscription are separate tracks.

May I use the numbers?

Yes, once published. Credit Aprex and link to this page. Every run gets a CSV file with all the numbers.

Which model should we choose?

Start with the task, not the model. Look at the score for the kind of work you will do, not just the total, and compare it with the cost. For simple tasks like sorting and extraction, a cheaper model may be good enough. Always test on a few of your own examples before deciding.

Can we use AI on customer data and personal data?

Yes, but settle a few things first: a data processing agreement with the provider, where the data is processed, whether the provider may train models on it, and that you send only what is needed. The GDPR also applies when an AI model processes the data.

Does the data stay in Norway or the EU?

It depends on the provider, product and agreement. Some offer processing in the EU or EEA, others do not. Check where data is stored and processed, and get it written into the agreement.

Is AI good enough in Norwegian and Nynorsk?

The barometer measures controlled structured tasks, not general fluency. The Nynorsk type has only controlled word-form and rule tasks. Check the scores once the first approved run is published, and test on your own texts.

What does running AI cost?

Through an API you pay per use, and prices vary a lot between models. Cost per 100 tasks will show the measured API price for tasks like those in the test. The subscription pilot reports no API cost; chat service licences have a fixed price per user. API and subscription are separate tracks and are never compared directly. Building and running the solution itself comes on top.

We already have ChatGPT or Copilot. Is that enough?

For employees who write, summarise and search, it often goes a long way. When a task must be done the same way many times, pull data from your own systems or run without someone chatting, you need an integration. Then the model behind it matters. The first pilot uses Codex and Claude Code with subscriptions. APIs and browser experiences are separate tracks; each measurement identifies product, model and version.

What does the EU AI Act mean for us?

The EU AI Act sets requirements by risk. Common uses such as customer replies, summaries and document extraction are usually not high risk, but there are transparency duties, for example that customers should know when they are talking to an AI.

Can we trust AI answers about laws and rules?

Not without checks. The barometer regulation type uses only supplied fictional rules, does not test legal expertise and is not legal advice. Use AI to find and summarise, and let a person check against the source before the answer is used.

How do we get started?

Pick one workflow that takes a lot of time and has clear rules. Collect ten to twenty real examples, test two or three models on them, and measure time and errors before and after. It is what we do in the barometer, only on your tasks.

Norwegian AI Barometer · Under development

How well does AI handle Norwegian work tasks?

The target is for Aprex to give 12 language models the same 48 work tasks with fixed answer keys in Bokmål and Nynorsk every week (12 public/12 calibration/24 holdout). We will measure how well they solve them, what API use costs and how long it takes. The first approved run has not happened. 12 public example tasks are available now.

See the results →How we test

Last updated: October 8, 2026 · v0.5-draft

Results

Latest measurement

Rank, score per work type, API cost and response time for every model. Rank shows observed scores, not proven superiority.

Under development. First approved measurement pending.

Twelve public example tasks are ready of 48 total (12 public/12 calibration/24 holdout). The test checks defined answers against a stored key. Numbers will appear after the first approved measurement. Free writing quality is outside the scope of this test.

48 tasks × 3 runs · 12 models

Results per model
Rank	Model	Total score	Extraction	Replies	Decisions	Rules	Classification	Nynorsk	API cost per 100 tasks	Response time, median
–	GPT-6 AstraOpenAI	Pending	–	–	–	–	–	–	–	–
–	Claude Opus 5.5Anthropic	Pending	–	–	–	–	–	–	–	–
–	Gemini 3.1 ProGoogle	Pending	–	–	–	–	–	–	–	–
–	Grok 4.7xAI	Pending	–	–	–	–	–	–	–	–
–	Mistral Medium 3.5Mistral	Pending	–	–	–	–	–	–	–	–
–	DeepSeek V4 ProDeepSeek	Pending	–	–	–	–	–	–	–	–
–	Qwen 3.8 MaxAlibaba	Pending	–	–	–	–	–	–	–	–
–	Muse Spark 1.3Meta	Pending	–	–	–	–	–	–	–	–
–	GPT-6 LunaOpenAIBudget	Pending	–	–	–	–	–	–	–	–
–	Claude Sonnet 5AnthropicBudget	Pending	–	–	–	–	–	–	–	–
–	Gemini 3.8 FlashGoogleBudget	Pending	–	–	–	–	–	–	–	–
–	Mistral Large 3MistralBudget	Pending	–	–	–	–	–	–	–	–

Scores run from 0 to 100, and the six task types count equally. The order shows observed scores, not proven superiority. Statistical comparisons on the same tasks will follow later. Cost is the API track's measured provider price, excluding VAT; subscriptions report no API cost.

Background

Why we built the barometer.

Businesses are adopting AI fast, but have little to go on when choosing a model for Norwegian tasks.

Existing tests are in English

The well-known language model tests mostly measure English, coding and maths. They say little about a Norwegian invoice, a customer reply or minutes in Nynorsk.

Businesses must choose a model

The choice is often made on gut feeling or marketing. Numbers on Norwegian work tasks give a better basis, including for cost.

We measure anyway

Aprex tests models when we build AI solutions. Now we do it systematically and publish the numbers with the method once the first approved run is ready.

What we measure

Six tightly scoped task types.

Eight tasks per type: 2 public, 2 calibration and 4 holdout. Difficulty levels are author-graded (2 easy/3 medium/3 hard per type), not empirically calibrated. All types count equally in the total score. Every task requires explicit JSON fields; no free prose is scored.