AI of the Week · Mon · Wed · Fri

The most important news from the AI field.

Summarized in Swedish, three times a week.

Sat May 17, 2026
The Year of Agents: Claude 4 runs 20-step sequences — qwen2.5:32b paired with GPT-4 locally

Claude 4 now regularly performs 15–20-step agent runs without human intervention. The category of possible automated tasks is dramatically broader than a year ago — for most Swedish SMEs, this means processes that required daily supervision can now be fully delegated to an AI agent.

qwen2.5:32b runs locally on an RTX 4090 or 5070 Ti and performs at the level of GPT-4 Turbo for summarization, classification, writing, coding, and analysis. Zero API costs, zero data leaves the network. For GDPR-sensitive industries — healthcare, law, accounting — it is an unparalleled game changer.

The Swedish Authority for Privacy Protection (IMY) has published preliminary guidance for AI agents under GDPR: if an agent makes decisions affecting physical persons, it requires a data protection impact assessment (DPIA). The more autonomous the agent is, the more documentation is required. Note: it is not a ban on autonomous agents — it is a requirement for traceability.

Read the whole issue →

Fri May 16, 2026
Grok Build in early beta — and OpenAI's legal battle against Apple

xAI opens Grok Build for an early developer beta — a platform for building applications with the Grok models. The API is similar to OpenAI's but includes access to xAI's real-time search on X (Twitter), providing a clearly differentiated offering for applications requiring fresh information.

OpenAI has initiated legal action against Apple regarding agreements surrounding AI integration in iOS 18. Details are confidential but signal an increasingly fragmented AI ecosystem. For Swedish companies building on Apple platforms, this could affect which AI services are built into future iOS versions.

Read the whole issue →

Tue May 13, 2026
The AI giants scale up — Google I/O, Microsoft Build, and what it means for SMEs

Google I/O and Microsoft Build coincided in timing and delivered an overload of AI announcements. Gemini 2.5 Pro in Google Workspace (Docs, Sheets, Gmail) and Copilot as a deeply integrated layer in Windows 11 and Microsoft 365 — that is the actual productization breaking through among broad business users during 2026.

The practical advice: if your company already pays for Microsoft 365 or Google Workspace — try the AI features now. The price is included. The barrier is zero. All that is required is that you and your team actually test them on real work tasks this week.

Read the whole issue →

Mon May 11, 2026
Meta spies on employees — OpenAI billions — and AI that halves coding costs

Meta has confirmed that the company collects internal communication data from employees to train its AI models. This raises questions that are extra sensitive in Europe — GDPR sets high requirements for employees giving informed consent, and it is unclear if Meta has this in place. For Swedish employers considering similar setups, the answer is simple: you cannot afford the legal risk.

OpenAI is now opening up for employees to sell shares on the secondary market, creating personal fortunes in the billions. It is a strategic move — it retains talent without new dilution and signals to the market that the company has sufficient cash. The valuation is solidified, and those who invested early in the AI industry are starting to see concrete results.

On the tools side: several independent studies now confirm that AI-assisted coding halves the cost of maintaining existing codebases. It's not about rewriting — but about understanding, documenting, and modifying old code. For Swedish tech companies with legacy systems, it is the most direct ROI calculation currently in the AI debate.

Fri May 9, 2026
OpenAI rolls out o3 globally — and lowers the price by 80%

OpenAI has begun rolling out o3 to all ChatGPT Plus subscribers globally, including Sweden — at no extra cost beyond the regular subscription of 220 SEK per month. The 80% price reduction for API access makes o3 available to an entirely new category of developers and companies that previously could not afford the model's capabilities.

o3 is OpenAI's strongest reasoning model and performs significantly better than GPT-4o on complex multi-step tasks, mathematics, and coding. For Swedish users, this means that a model that previously required separate API costs is now directly available in the familiar ChatGPT interface.

The practical advice: test o3 on tasks where you usually need to follow up on the AI response several times — complex calculations, legal summaries, or code with many dependencies. That is where the difference from previous models is most noticeable.

Wed May 7, 2026
Google Gemini 2.5 Pro beats Claude in coding — but not in Swedish

Google's Gemini 2.5 Pro now takes first place on several coding benchmarks, including HumanEval and SWE-bench, and the margin against Claude 3.7 Sonnet is measurable. The model handles long code contexts exceptionally well and is quick at refactoring large codebases.

Swedish tests show a clear weakness: Gemini 2.5 Pro tends to mix in English terms in otherwise Swedish responses, and the idiomatic precision in formal business texts is noticeably lower than Claude's. For a Swedish entrepreneur who wants AI to help with customer communication, quotes, and internal materials, it's this difference that makes the decision.

Conclusion: if your primary use case is code — and the code is in English — Gemini 2.5 Pro is worth testing. However, if you use mixed cases with Swedish texts as core content, Claude still holds the superiority where it matters for Nordic businesses.

Mon May 5, 2026
Anthropic opens Claude for enterprise automation — new API tier

Anthropic launches a new enterprise API tier called "Claude for Work" specifically aimed at medium and large companies looking to build internal automation flows. The tier offers increased rate limits, dedicated capacity, and a new system for Organization management that allows IT departments to control access per team and project.

What sets it apart from existing API offerings is the built-in auditing log — every call is logged with timestamps and user IDs, making it easier to comply with requirements under the EU AI Act regarding traceability for systems that make or support decisions. For Swedish SMEs handling customer data, this is a concrete relief.

The pricing model is volume-based with a base price of around 500 dollars per month. For a company running hundreds of internal automation calls daily, it's reasonable — but it also means the threshold is too high for individual consultants and micro-businesses. However, it opens up opportunities for resellers and system integrators to build Claude-driven products for the Swedish market.

Next issue · Sunday

Polaris in your inbox —
Monday, Wednesday, and Friday — always free.

The most important things in AI, for everyone who wants to learn to master it. No noise. Just what actually matters.

Free forever · cancel anytime