Which AI should you use in autumn 2026? ChatGPT, Claude, Gemini, Grok and the open models, compared for ordinary people and small businesses
GPT-6, Claude 5.5, Gemini 4 Argon, Grok 4.7 and a flood of open models: the choice has never been bigger – or the gaps at the top smaller. Here is what actually separates them in October 2026, what they cost in Norway, who uses your data, and what to pick for what.
Seven in ten Norwegians aged 16 to 79 used generative AI in the past three months, according to Statistics Norway's September figures – up from five in ten last year – and nearly every second company with more than ten employees uses AI. At the same time, autumn has delivered a new generation of models within a single month: GPT-6 from OpenAI, the Claude 5.5 line from Anthropic, Gemini 4 Argon from Google, Grok 4.7 from xAI, Mistral Large 4 from Europe and a flood of open models from China. The question we hear most often is simple: which one should I use? The answer in October 2026 is less dramatic than the marketing: at the top, the models have become so alike that the choice comes down to price, privacy, language and the ecosystem you already live in.
What has happened since the summer
- OpenAI launched GPT-6 Astra on 3 September as its most capable model, followed by the budget GPT-6 Sol and Luna on 22 September and an updated GPT-6.1 Sol. Since 7 October "GPT-6 Sol" is the default in ChatGPT Plus, and free users got GPT-6 Luna the day after.
- Anthropic released Claude Fable 5.1 on 1 September, Claude Opus 5.5 on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October. Opus 5.5 is priced at less than half of Fable, and Anthropic says it performs at Fable's level "on most work". This summer's Claude Opus 5 is already "legacy".
- Google launched Gemini 3.8 Flash on 2 September and announced Gemini 4 Argon on 30 September – but Argon is so far available only to selected cyber-defence organisations and US government review, then paying API customers and Ultra subscribers. In the Gemini app, 3.1 Pro is still the "Pro" model.
- xAI has had Grok 4.7 in its API since September, but not yet in the consumer apps.
- Mistral unveiled Mistral Large 4 on 6 October – around a trillion parameters, 160 languages including every official EU language, served from European data centres – with open weights promised by the end of the month.
- Open models: Kimi K3 (which we covered in July), Xiaomi's MiMo-V2.6-Pro, GLM-5.3, DeepSeek V4.1, Qwen 3.8 and Google's Gemma 4. And Norway's own Borealis 2: the National Library's model built on Gemma 4, further trained on Norwegian newspapers under a collective licence with Kopinor that the state pays 45 million kroner a year for, released as a preview on 28 September.
Who is best? The answer is "almost all of them"
The two most-used yardsticks say the same thing. On the Artificial Analysis Intelligence Index, which on 10 October ranks 205 models, the top eleven sit within six points: Claude Opus 5.5 at 58, Sonnet 5.5 at 56, Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon all at 53, and GPT-6.1 Sol at 52. The cost per task, by contrast, varies eightfold – from 72 cents for GPT-6.1 Sol to nearly six dollars for Opus 5.5 at maximum effort. On LMArena, where humans vote blind, Gemini 4 Argon led provisionally on 8 October – on fewer than 5,000 votes and with a lead inside the margin of error – ahead of Claude Opus 5.5, and no GPT-6 model was in the top ten.
On coding, the vendors argue over their own tests: Google says Argon leads DeepSWE, Anthropic that Opus 5.5 leads Terminal-Bench and OSWorld, xAI that Grok 4.7 is second best at coding. Two research papers this year point at the same problem: small gaps at the top of such leaderboards are not statistically meaningful, and the vendors choose the setup, effort level and number of attempts themselves. The practical conclusion: for everyday writing, summarising, translating and analysis, every frontier model is good enough. Choose on the other criteria.
What it costs in Norway
- ChatGPT: free with GPT-6 Luna. Go costs NOK 99 a month and Plus NOK 249 in the App Store (prices include VAT and Apple's margin; on the web Plus is USD 20 plus VAT). Pro costs USD 100 or 200.
- Claude: free with Sonnet 5.5 and Haiku 5.5 – that is, the second-best model on the Artificial Analysis index, without paying. Pro costs USD 20 a month (17 on annual billing) plus VAT and unlocks Opus 5.5 and Fable 5.1; Max from USD 100.
- Google: free with Gemini 3.5 Flash-Lite, AI Plus NOK 64, AI Pro NOK 259 with 3.8 Flash, 3.1 Pro and 5 TB of storage, AI Ultra from NOK 1,059 – in Norwegian kroner, straight from Google.
- Microsoft: Microsoft 365 Personal with Copilot in the Office apps NOK 121 a month, Premium with "extensive AI usage" NOK 259. For businesses, Microsoft 365 Copilot costs NOK 213.50 per user per month excluding VAT on annual billing – NOK 183 on promotion until the end of 2026 for existing customers.
- Grok: free on X and grok.com with limits; paid plans are priced in dollars, and xAI publishes no Norwegian prices.
- Mistral Vibe (formerly Le Chat): free, Pro USD 14.99.
Who writes the best Norwegian?
The differences are smaller than you might think. At the Klarspråk plain-language conference in Bergen in March, the tests concluded that the big models deliver roughly equally in Norwegian, and that Mistral and the Norwegian NorMistral were as good as ChatGPT and Copilot in both Bokmål and Nynorsk – "as long as they are used with awareness and language expertise". The Language Council of Norway's test found that every model writes worse Nynorsk than Bokmål, mainly on grammar. According to an independent test in July, the Norwegian Borealis model writes native-sounding Bokmål, better Nynorsk and better terminology than the big ones, but loses on reasoning and the structure of long documents. Gemini is missing from the Norwegian tests. Our experience from daily use: Claude and ChatGPT for long Norwegian texts, Borealis locally when Nynorsk, terminology or privacy weigh most.
Who uses your data?
This is the criterion people skip, and they should not:
- ChatGPT uses conversations from free and personal plans for training unless you turn off "Improve the model for everyone" under data controls. Temporary chats are never used. Business, Enterprise and the API are not used.
- Claude has since August 2025 required consumer users to make an active choice; say yes and conversations are kept for five years, say no and for 30 days. Team, Enterprise and the API are never used.
- Gemini trains on conversations when "Keep Activity" is on, and a sample is read by human reviewers – and kept for up to three years even if you delete it. With the setting off, chats are deleted after 72 hours.
- Meta AI has, in the EU, used adults' public posts and interactions with the AI since April 2025; you have to object actively through a form. Private messages are not used.
The Norwegian Data Protection Authority's advice to employers is to write guidelines and be transparent about what is collected. The practical rule every lawyer repeats: do not paste personal data about customers or staff into a free chatbot. Use the business tiers with a data processing agreement and storage in the EEA – or a model you run yourself.
The rules: the EU is moving, Norway lags
The EU AI Act has applied to general-purpose models since August 2025, and since 2 August 2026 the transparency duties apply – including that AI-generated content must be labelled. The so-called Digital Omnibus, adopted this summer, postponed the high-risk rules to December 2027 and August 2028. Norway is not there yet: the regulation has not been incorporated into the EEA Agreement, a new consultation is announced for this autumn, and the government is aiming for a bill to parliament in spring 2027. Until then, the GDPR and the Data Protection Authority apply.
Run it yourself – or in Europe
More and more people can. Rule of thumb: models of 2–8 billion parameters run on a phone or any laptop; 12–31 billion (Gemma 4, Qwen 3.8-27B, Mistral Small, Borealis 2) need 24–32 GB of memory or one gaming GPU; the large open models of 750 billion to 2.8 trillion parameters (DeepSeek V4, GLM-5.3, Kimi K3 at 1.4 terabytes) need a rack of data-centre GPUs – or an API. For companies that want their data in Europe there are now real options: Mistral lets you choose EU servers, France's Scaleway serves open models from 15 cents per million tokens, and Telenor AI Factory runs more than 190 open models from Norwegian data centres. We have written before about why electricity makes Norway a natural home for this.
So: what should you choose?
- Writing, thinking and reading in Norwegian: Claude Pro or ChatGPT Plus – try both for free first. If you need Nynorsk or technical vocabulary, Borealis 2 locally.
- Free: Claude Free currently gives the most model for nothing, ChatGPT Free is the broadest, Google AI Free includes Deep Research.
- You live in Google Workspace: Google AI Pro at NOK 259.
- Your company lives in Microsoft 365: Copilot at NOK 213.50 per user – NOK 183 on promotion until year end.
- Coding: Claude Opus 5.5 or Sonnet 5.5 lead the lists; Grok 4.7 is good value; GitHub Copilot lets you switch between them.
- Research and search: Perplexity for sources, Gemini Deep Research (free), ChatGPT search for breadth. Remember that Google search itself has become AI.
- Images and video: gpt-image-2.5 in paid ChatGPT and Google's Nano Banana for images; Veo 3.1 for video; Midjourney for art.
- Personal data: a business tier with a data processing agreement and EU storage, or an open model at a European provider.
Frequently asked questions
What is the difference between ChatGPT and Claude?
In performance: little. ChatGPT has the most features, apps and users; Claude scores highest on the independent indexes and offers a stronger free model. Privacy is a choice with both.
Which AI is free?
All the big ones have free tiers: ChatGPT (GPT-6 Luna), Claude (Sonnet 5.5), Gemini (3.5 Flash-Lite), Grok, Mistral Vibe and Meta AI. Borealis, Gemma and Qwen you can download and run yourself.
What does ChatGPT Plus cost in Norway?
NOK 249 a month via the App Store, or USD 20 plus VAT on the web – around NOK 250 depending on the exchange rate.
Can I use ChatGPT at work with personal data?
Not the free version. Use a business tier with a data processing agreement, or a model running at a Norwegian or European provider, and follow your employer's guidelines.
When does the AI Act take effect in Norway?
Autumn 2027 at the earliest. The government plans a bill in spring 2027 after a new consultation this autumn.
Do the AI models write Nynorsk?
Yes, but worse than Bokmål. Borealis and NorMistral do best in the Norwegian tests; ChatGPT and Copilot are usable with proofreading.
