Norwegian AI Barometer · Calibration 0.1

How well does AI handle Norwegian work tasks?

Aprex tests AI on Norwegian work tasks with a fixed answer key. Explore the results and see how we check them.

Results updated

Results · Calibration

Latest results

Advanced calibration 0.1 · 12 synthetic Bokmål tasks · one repetition per model. Eight HIGH routes (96 responses from two runs), plus a separate Qwen xhigh test.

Models at HIGH

Alphabetical · 12 tasks per model
  • Claude Haiku 5.5HIGH11/12fully correctDetails

    Correct fields: 290/292. diagnostics · no partial credit.

    Stock and orders
    4/4 tasks80/80 fields
    Invoice reconciliation
    4/4 tasks104/104 fields
    Case flow and deadlines
    3/4 tasks106/108 fields
    Source and run setup
    model ID
    anthropic/claude-haiku-5-5
    run
    advanced-v01-2026-10-10-haiku · 2026-10-10
    Provider route
    Anthropic
    harness
    Codex native subagents
    commit
    5f55d82430baa317bbd8dff30829b2d799f75b6b
    definition
    e3f59476abe8455fd4e519fb3dea247270af42f7f0dddfaac4f7b861cd6309eb
    Manifest / scores
    0a084315b28d5438ad2a2a3daf406048c8138dc0ce5c4bd763547929327e7b6e
    48f5b0998db0aef3517533240a4a93946c71ef58facb8b24599b94d7ee9468fd
  • Claude Opus 5.5HIGH11/12fully correctDetails

    Correct fields: 291/292. diagnostics · no partial credit.

    Stock and orders
    4/4 tasks80/80 fields
    Invoice reconciliation
    3/4 tasks103/104 fields
    Case flow and deadlines
    4/4 tasks108/108 fields
    Source and run setup
    model ID
    anthropic/claude-opus-5-5
    run
    advanced-v01-2026-10-10-opus · 2026-10-10
    Provider route
    Anthropic
    harness
    Codex native subagents
    commit
    5f55d82430baa317bbd8dff30829b2d799f75b6b
    definition
    722c15b7dd27939c5b298d6efa6c7f49eb9a612ef70ab1fe3e726e36b7994678
    Manifest / scores
    a2f2e872ee49b14d57e1c11b3354dbbd5e3f81271580302178fb2921c34e4fc7
    16f082c0966955b427aad0d930c8c9b6e668129aa5c5d616a15354d183a30e6e
  • DeepSeek V4.1 FlashHIGH9/12fully correctDetails

    Correct fields: 288/292. diagnostics · no partial credit.

    Stock and orders
    4/4 tasks80/80 fields
    Invoice reconciliation
    4/4 tasks104/104 fields
    Case flow and deadlines
    1/4 tasks104/108 fields
    Source and run setup
    model ID
    opencode-go/deepseek-v4.1-flash
    run
    advanced-v01-2026-10-10-deepseek · 2026-10-10
    Provider route
    OpenCode Go
    harness
    Codex native subagents
    commit
    5f55d82430baa317bbd8dff30829b2d799f75b6b
    definition
    90b1a48c6548e487b81a1c802ad7c1ab6ed0d206edb8f0e58326b1f1f7a021ee
    Manifest / scores
    7b079382e576759867a991e77bfc97c9e78e6be58b11553481d7be59027a4cae
    19c2f610da3d9b3c9b79b5e6cc46884401f8be4a347d230154ca3384ac142306
  • GPT-6 AstraHIGH12/12fully correctDetails

    Correct fields: 292/292. diagnostics · no partial credit.

    Stock and orders
    4/4 tasks80/80 fields
    Invoice reconciliation
    4/4 tasks104/104 fields
    Case flow and deadlines
    4/4 tasks108/108 fields
    Source and run setup
    model ID
    openai/gpt-6-astra
    run
    advanced-v01-high-extension-2026-10-10-astra · 2026-10-10
    Provider route
    OpenAI
    harness
    Terminal Codex CLI 0.155.1 --ignore-user-config, ChatGPT login, read-only sandbox, empty workspace
    commit
    a230b80a22b962b1f69f39170b22950db657986d
    definition
    fccb8e875ae7fbc0546bd740eddc9c27a4a830bdb11414baed5964c34e1a311a
    Manifest / scores
    a6e1ea40acd26fe488d294a8aaf9385bf20b4b70ad8a6cc0daaa3c4bfdefa76b
    50b516678c5c3db2397f923ade83ea3674956282fb6934b4751432fd0581fbad
  • GPT-6 LunaHIGH4/12fully correctDetails

    Correct fields: 270/292. diagnostics · no partial credit.

    Stock and orders
    3/4 tasks79/80 fields
    Invoice reconciliation
    1/4 tasks96/104 fields
    Case flow and deadlines
    0/4 tasks95/108 fields
    Source and run setup
    model ID
    codex/gpt-6-luna
    run
    advanced-v01-2026-10-10-luna · 2026-10-10
    Provider route
    OpenAI
    harness
    Codex native subagents
    commit
    5f55d82430baa317bbd8dff30829b2d799f75b6b
    definition
    544706915eea14f643d4bb929df1b47748519423db5f3c4ac5243401ff5a2174
    Manifest / scores
    71e4e55a203776ffd8ce4d0d2708d6a3e2cfd21958217bcf7e8ff31fae7baa9f
    078d9d0f541daad196d654abed9124978f9618340c34aadac5eeb3424807b87f
  • GPT-6.1 SolHIGH12/12fully correctDetails

    Correct fields: 292/292. diagnostics · no partial credit.

    Stock and orders
    4/4 tasks80/80 fields
    Invoice reconciliation
    4/4 tasks104/104 fields
    Case flow and deadlines
    4/4 tasks108/108 fields
    Source and run setup
    model ID
    openai/gpt-6.1-sol
    run
    advanced-v01-high-extension-2026-10-10-sol · 2026-10-10
    Provider route
    OpenAI
    harness
    Bundled app Codex CLI 0.162.0-alpha.17.2, user config loaded, ChatGPT login, read-only sandbox, empty workspace
    commit
    a230b80a22b962b1f69f39170b22950db657986d
    definition
    9ffb1f7958ea5f323cb0e4c340f6ccb3d3012a4d5230ab4cc36a0dcb9dee233b
    Manifest / scores
    c555ade670f7d768f32e1ab7835411b329d6e61690bb3ed7649f962bb97b0a06
    2e96682bb57236d544b6ed2027dfbc0d3399c39383f41bbf87aed20b185a5d98
  • Gemini 3.8 FlashHIGH12/12fully correctDetails

    Correct fields: 292/292. diagnostics · no partial credit.

    Stock and orders
    4/4 tasks80/80 fields
    Invoice reconciliation
    4/4 tasks104/104 fields
    Case flow and deadlines
    4/4 tasks108/108 fields
    Source and run setup
    model ID
    google/gemini-3.8-flash
    run
    advanced-v01-high-extension-2026-10-10-gemini · 2026-10-10
    Provider route
    Antigravity UI
    harness
    Antigravity native UI, No Project, new conversation per task
    commit
    a230b80a22b962b1f69f39170b22950db657986d
    definition
    bc4762926981f3674f92d986415fef29c50d1f4b550a0e2c48f3efe929f413a8
    Manifest / scores
    172889dcbcb3d48f53b1c8f90d9357bcd03d2ffc2f64c604375880822b5d4004
    d31d7434cf37fc33733c85fa7ea62ffabd6a9214a6088b457327287717515a59
  • Muse Spark 1.3HIGH10/12fully correctDetails

    Correct fields: 289/292. diagnostics · no partial credit.

    Stock and orders
    4/4 tasks80/80 fields
    Invoice reconciliation
    2/4 tasks101/104 fields
    Case flow and deadlines
    4/4 tasks108/108 fields
    Source and run setup
    model ID
    meta-muse/muse-spark-1.3-contributor
    run
    advanced-v01-high-extension-2026-10-10-muse · 2026-10-10
    Provider route
    Meta Muse
    harness
    Codex subagent V1
    commit
    a230b80a22b962b1f69f39170b22950db657986d
    definition
    199d5ef3e6d9e75a45e6855d3dc090f390771dc905f61d7227da6a90dce56e33
    Manifest / scores
    1473ea0c803278eddbff91a53015d0eb6966b8fa36ecd514ad9574747aff456b
    d75dc632ce7a7aeab8ddc6da543b5390c1bfd4b6e3dba0074b0ff1af7e4b3281
Compare task types
Latest selection per task family
ModelStock and orders4 tasks · 80 fieldsInvoice reconciliation4 tasks · 104 fieldsCase flow and deadlines4 tasks · 108 fields
Claude Haiku 5.54/480/80 fields4/4104/104 fields3/4106/108 fields
Claude Opus 5.54/480/80 fields3/4103/104 fields4/4108/108 fields
DeepSeek V4.1 Flash4/480/80 fields4/4104/104 fields1/4104/108 fields
GPT-6 Astra4/480/80 fields4/4104/104 fields4/4108/108 fields
GPT-6 Luna3/479/80 fields1/496/104 fields0/495/108 fields
GPT-6.1 Sol4/480/80 fields4/4104/104 fields4/4108/108 fields
Gemini 3.8 Flash4/480/80 fields4/4104/104 fields4/4108/108 fields
Muse Spark 1.34/480/80 fields2/4101/104 fields4/4108/108 fields

Separate test: xhigh

Separate xhigh run (excluded from HIGH numbers)

  • Qwen 3.8 MaxXHIGH12/12fully correctDetails

    Correct fields: 292/292. diagnostics · no partial credit.

    Stock and orders
    4/4 tasks80/80 fields
    Invoice reconciliation
    4/4 tasks104/104 fields
    Case flow and deadlines
    4/4 tasks108/108 fields
    Source and run setup
    model ID
    opencode-go/qwen3.8-max
    run
    qwen-xhigh-isolated-2026-10-10 · 2026-10-10
    Provider route
    OpenCode Go
    harness
    OpenCode CLI 1.18.33 --pure, OpenCode Go subscription, explicit isolated cwd outside repo, all tools denied
    commit
    a230b80a22b962b1f69f39170b22950db657986d
    definition
    75d73dca32bf2f5c9f4c94f4100e6a2787a5ab43fd67218d8dab3242f05a0ea4
    Manifest / scores
    35f4c3da16476b15e1f4e73ab5407fea6d72175f43a3e71e7663742d64eb311a
    4e2c0130fdf98dcc45858540f97bfa6e053b544b03df9f335b163e0117167d80
    audit
    7a16849f6dff0cf0ea1af7ec7029b8919e34bdaf137492d060a6b56e8f663c54
    Check
    Audited separate xhigh run on the same frozen tasks; not comparable to HIGH as an equal-compute test. One repetition, reused calibration, no ranking. Both earlier batches excluded whole for harness context before scoring; raw outputs preserved.
    Run context
    OpenCode CLI isolated cwd verified in each exported session and assistant path; isolated config, neutral local AGENTS, Claude compatibility disabled, plugin loading disabled, all tools denied; exact prompt inside documented quote wrapper; watchdog 900 seconds. Local client-declared identity, not signed provider attestation.
Documentation, downloads and corrections
Version
Calibration report · advanced-0.1-utkast
Run
10 October 2026
Published
10 October 2026
Selection
8 HIGH routes from two October 10 runs · not one 96-answer run
Scope
12 tasks × 1 run · 292 fields per model · 8 routes
Responses
96 of 96 checked
Checks
0 missing · 0 technical errors · 0 protocol-rejected · 0 tool calls
Traceability
snapshot 5132dfd81cbb · scorer 4f26f9806b58 · audit ca60602200a7

Corrections October 10, 2026

  • Sol was first rejected in terminal Codex 0.155.1 with --ignore-user-config before answering. The bundled app Codex 0.162.0-alpha.17.2 with user config then completed all 12 tasks at HIGH, transcript-audited 12/12 and 292/292. The first attempt is preserved, not scored and not hidden. Two settings differed; the single cause is not isolated. No general block.
  • Gemini logs verified on October 10: the same 12 conversations were checked against complete local transcripts with 12/12 exact prompt matches, 12/12 answer matches, logged Gemini 3.8 Flash (High) and 0 recorded tool calls – still 12/12 and 292/292. The evidence is local app logs, not signed provider attestation; the model name is client-declared; the app wrapper and system instructions differ from the other runs. Timestamps are not used as response-time measurements.
  • October 10: two Qwen batches were excluded before scoring and are preserved, not scored – first 9 started (6 completed, 1 database stop, 2 interrupted, 3 not started) because the workspace sat under the repo and AGENTS files could have contaminated the context; then 11 started (7 completed, 1 watchdog stop, 3 interrupted, 1 not started) because an inherited PWD registered the repo as the working directory. Combined: 20 started and 13 excluded answers.

Separate xhigh run (excluded from HIGH numbers)

  • 12 responses · 0 missing · 0 technical errors · 0 protocol-rejected · 0 tool calls (separate run)
  • Qwen 3.8 Max at xhigh appears here as a separate run, outside the HIGH comparison. xhigh numbers are never pooled with HIGH numbers.
  • Separate clean run outside the repo: verified working directory, no plugins, all tools denied and a 900-second watchdog. The task text sat in literal quotes around the frozen prompt; the frozen text inside is exact.

Scope and caveats

  • 12 synthetic Bokmål tasks with 292 answer fields per model, in three families of 4 tasks each. Each task ran once per route. “Whole tasks” is primary (all-or-nothing); correct fields are diagnostics only with no partial credit.
  • 96 of 96 responses were checked against the run plan and logs: 0 missing, 0 technical errors, 0 protocol-rejected answers and 0 tool calls. The latest selection combines two October 10 runs with different harnesses; it is not one joint 96-answer run. The 12 Gemini responses were checked against local transcripts and are not signed provider attestation.
  • Different harness and background context: Codex subagents, Codex CLI, the bundled app CLI and Antigravity UI have different product instructions. HIGH does not mean equal compute across models.
  • For the HIGH routes, tools were prohibited in the task prompt and audited in the logs; they were not technically removed from the candidate environment. The separate Qwen run denies all tools technically.
  • Model names are the client’s declared identity, not independently confirmed by the underlying provider.
  • Muse helped write the task generators. Fresh candidate contexts and separate checks reduce leakage risk but do not remove author bias.
  • Sonnet was not in the available subagent catalog. No Grok routes were run. Qwen 3.8 Max ran at xhigh as a separate run (12 responses), outside the HIGH comparison.
  • The 24 reserved holdout tasks in the core protocol remain untouched. This calibration is a separate development signal — no full run and no ranking — and is never combined with holdout into a claimed measurement on unseen tasks.
  • Before any public ranking, new pre-frozen tasks and repeats must show whether the differences hold. Alphabetical order; no winner.

How to cite

Aprex AS, Norwegian AI Barometer, advanced calibration 0.1, current audited selection October 10, 2026 (advanced-v01-audited-selection-2026-10-10). Link to this page and state the CSV/JSON date, the selection and the corrections.

The current test

Three types of work

Four tasks in each category. All documents and rules are fictional. The test measures precise answers, not free writing or legal expertise.

01

Stock and orders

Allocate limited stock and resolve competing orders using the supplied rules.

02

Invoice reconciliation

Match invoices, payments and adjustments across several documents.

03

Case flow and deadlines

Apply exceptions and calculate deadlines from a fictional case history.

Public examples from the planned core test →

Method

How we test.

The published numbers are a calibration. The first full evaluation on reserved tasks is still ahead.

  1. 01

    A fixed answer key

    The program compares the response with a key computed from synthetic data. There is no AI judge.

  2. 02

    All fields must match

    A task is correct only when every required field is correct. Field scores help explain errors and give no partial credit.

  3. 03

    Logs checked before publication

    The prompt, response, model setting and tool history are checked. We publish the tool and reasoning level with each result.

Read the full method: scoring, uncertainty and data format →

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. The latest HIGH selection from October 10 is published with Gemini logs verified; earlier numbers are preserved as history. The first approved full run is pending.

Next phase: planned task types, models and public examples →

Changelog

Results and history

Previous reports stay available. Different test versions must not be read as a performance trend.

Earlier calibration (history, unchanged)

Earlier report versions are preserved.

Changes and corrections
  • October 10: the separate Qwen 3.8 Max xhigh run is audited (12/12 tasks, 292/292 fields, 12 responses with zero findings) and published with its own files outside the HIGH comparison. Today's selection is 8 HIGH routes with 96 responses plus separate xhigh with 12 responses. The exclusion of two earlier Qwen batches (20 started, 13 answers) is documented as a dated correction.
  • Correction October 10: Gemini 3.8 Flash HIGH logs are now verified against complete local transcripts (12/12 exact prompt and answer matches, logged Gemini 3.8 Flash High, 0 recorded tool calls) and join the current audited selection: 8 HIGH routes and 96 responses, same 12 tasks and 292 fields. The transcript evidence is local, not signed provider attestation; the model name is client-declared; the app wrapper differs from other runs; one repeat, reused calibration tasks, no ranking. The 7-route October 10 selection, the October 9 report and Muse MAX are preserved as history.
  • October 10: two Qwen batches (20 started attempts and 13 completed answers combined) were excluded before scoring and are preserved, not scored – first possible AGENTS context contamination, then an inherited PWD registering the repo as the working directory. This was the status before the clean run. The completed and audited Qwen run is documented in the entry above and shown separately from the HIGH comparison.
  • Published the latest HIGH selection from October 10: 7 audited routes, 84 responses, same 12 tasks and 292 fields. Provisional Gemini result shown separately and excluded. Sol correction from October 10 included. The October 9 report and Muse MAX are preserved as history.
  • Published advanced calibration 0.1 from 9 October: 12 tasks, 292 answer fields and four model routes. Reports fully correct tasks and field diagnostics, with CSV and method limitations. This is a calibration report; the holdout evaluation has not been run.
Changes and corrections
  • Protocol 0.5-draft: six hard calibration tasks have been revised after five model routes achieved full scores in the first 0.4 calibration. The tasks require several connected source selections, calculations and exceptions. Six easy control tasks, twelve public examples and 24 reserved final-evaluation tasks are retained. Answer keys and scoring rules are frozen before a new run. Codex-subagent calibration has separately documented isolation and transcript checks. Calibration is method development, not a public final ranking. No approved results.
Changes and corrections
  • Protocol 0.4-draft: 48 tasks (12 public / 12 calibration / 24 holdout), eight per type with 2 easy, 3 medium and 3 hard as author-graded draft without empirical calibration. 12 public example tasks, two per type. Strict 100/0 stays the ranking score; share of correct fields is diagnostics only, null only where the key explicitly allows it, and missing answers and interrupted calls are listed separately without 0. Task prompts carry rules for unknown, missing or conflicting information. No AI judge. Subscription runner is Codex and Claude Code only with no API fallback, using --split and taskSplit (default public). Primary comparison is the 24 reserved holdout tasks; the publication gate still requires full 48-task coverage, so the release protocol must be updated before future publication. Freeze before models are called, version bump on change, used holdout is retired with a versioned new holdout, no edits because a model disagrees. Empirical calibration on the calibration split and final publication are pending. No approved results. First approved measurement pending.
Changes and corrections
  • Protocol 0.3-draft: strict structured tasks with 100/0 field scoring for all six types, no checklists or judges in the score. Regulation uses fictional rules, Nynorsk corrects only listed forms. 6 example tasks available of 48 planned. First measurement pending.
  • The barometer launched with test protocol 0.2, twelve models and six public example tasks. The first measurement has not been run yet.
For businesses

What Norwegian businesses ask.

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. Read more →

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. Read more →

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 full run is published — the calibration above is only a development signal on 12 Bokmål tasks — 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. The published calibration contains no cost or response-time measurements. Building and running the solution itself comes on top. Read more →

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. Read more →

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. Read more →

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. Read more →

FAQ

Questions and answers

What is the Norwegian AI Barometer?

A barometer under development by Aprex AS (0.5-draft) for tightly scoped Norwegian work tasks with verifiable answer keys in Bokmål and Nynorsk: 48 tasks (12 public/12 calibration/24 holdout). Difficulty levels are author-graded, not empirically calibrated. The first approved full run is not published. The latest HIGH selection from October 10 (12 synthetic Bokmål tasks, 8 routes, 96 responses, plus separate Qwen xhigh with 12 responses) is published as a separate development signal. Full results will show scores per split with holdout-24 as the primary comparison, API cost and response time per model.

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. The catalog has no results; the latest selection covers only the 8 actually run HIGH routes. Changes to the list are recorded in the changelog.

How often is the barometer updated?

The target is every week. No approved full runs have been published yet; advanced calibration 0.1 is published as a separate development signal. Full 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. The calibration uses the same 100/0 rule.

Could the models have seen the tasks before?

The 12 example tasks on the method 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. The 12 advanced calibration tasks have been used for development and are no untouched final test; the 24 holdout tasks remain untouched.

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. The calibration reports no cost or response time.

May I use the numbers?

Yes. The calibration numbers can be used now. Credit Aprex and link to this page. Every run gets a CSV file with all the numbers.

Limits

What the numbers do not say

The latest selection covers 12 synthetic Bokmål tasks and 8 HIGH routes from two runs with different harnesses, plus separate Qwen xhigh with 12 responses, presented alphabetically with no ranking and no claim of general superiority. Full results will apply to these 48 controlled tasks and to the product and model versions available in the week they were measured. Difficulty levels are author-graded, not empirically calibrated. Free writing, tone and fluency are not measured. Aprex itself uses models from several of the providers. The providers have no say in the method and do not see results before they are published.

Which model fits your tasks?

The barometer shows an average on controlled tasks. Tell us what you want to use AI for, and we will help you choose and test the right model on your own tasks.

Tell us what you want to test →