AI in klare taal edition
OpenAI is holding back its strongest model. Meanwhile, cheaper AI work is getting closer.
OpenAI opts for a cheaper model instead of its top model
OpenAI says GPT-6.1 Sol costs a fifth of the price of Astra. And Sol can be used for work.
OpenAI is not releasing its latest top model, GPT-6.1 Astra. The company cites safety reasons. Instead, GPT-6.1 Sol will be available. OpenAI says Sol is almost as good at three things: coding, using a computer and professional work. Coding means writing software. Yet Sol costs only a fifth of Astra’s price. Through the API, Sol costs about 2 dollars per million input tokens. Output costs 10 dollars per million tokens. The API is a link that lets companies use the model in their own software. A token is a small piece of text that an AI processes. Sol works through the API, ChatGPT Work and Codex. For companies, this can greatly reduce the cost per task. The price depends on how much text the model reads and sends back. If a company uses Sol for many tasks, it may pay less than it would for Astra. That is only true if Sol works well enough for those tasks. “Almost as good” applies only to the tasks OpenAI mentions. So the lower price does not prove that Sol is as good as Astra for every kind of work. The source also does not say how much an average office task costs. There is no comparison test for a regular email or a short summary. OpenAI is holding Astra back for safety reasons. But the source does not explain which risks led to that decision. For employees, the main thing is which model their employer allows, and for what. Ask your administrator which model you use. Also ask whether the cost per task varies. And ask how that choice is checked. Do this before handing sensitive or important tasks to an AI model.
Background for subscribersA lot of this week’s news is about choices made elsewhere. Which model does your employer choose? Which saved instructions does a tool keep? What agreements do companies make themselves? You can still take action. Ask your administrator which AI model you use and whether the cost per task varies. Review saved instructions before using them widely, such as a Gemini skill. Check an AI answer against a second source. Keep track of where protected material came from. If you publish an AI image on behalf of your organisation, check the facts and get permission first. And ask what AI training is available where you work. A voluntary agreement or a lower price does not mean a tool is good enough for your work.
OpenAI DevDay: ‘dots’ (AI helpers that are always on), ChatGPT Space and 20+ new features
On 29 September, OpenAI showed more than 20 products at DevDay. The main idea is ‘dots’. These are agents in ChatGPT that are always on. Agents are AI helpers that carry out tasks for you. They reschedule appointments, book flights or change code on your behalf. ChatGPT Space is also coming. It is a shared workspace. Teams, ChatGPT and dots work together on projects there. There is also a Decisions API. An API is a link to other software. It lets companies hand small, repeated decisions over to AI. Plugins (add-ons) work as full apps in ChatGPT. The launch comes at a time of debate. People are asking how well AI that works on its own can be controlled.
Background for subscribersGemini lets you reuse saved instructions
Google is making skills available worldwide in Gemini chat. A skill is a saved set of instructions for a task that comes up again and again. You can ask Gemini to create one from a conversation. After that, Gemini starts it automatically when you type a matching request. Google says skills replace the earlier Gems. At work, this can be useful for recurring texts. For example, reports, emails or quotes. You do not have to repeat your instructions each time. But a skill that starts automatically can shape the result. This can happen even if you do not notice which skill has started. An error in the saved approach can also carry over to several requests. The source does not say what checks Gemini makes before starting a skill. It also gives no rules for access or subscription terms. So treat a skill as a standing work instruction that you need to check. It does not guarantee that every result is correct. Review one skill this week. Adjust the instructions to match the way you work. Go over the text with your team before using it widely.
Background for subscribersVoluntary agreement promises checks, but no one enforces them
Trump and six AI companies signed a voluntary safety agreement on 29 September. They are OpenAI, Google, Anthropic, Meta, Nvidia and xAI. They promise internal checks on their most powerful models. They also promise external reviewers and a safety committee on their boards. The agreement is not part of the law. There are no penalties and no deadlines. And the companies choose their own reviewers. An external check sounds independent. But how strict it is depends on how companies put the agreement into practice. This is called self-regulation: companies decide for themselves which checks they carry out. That is different from a binding rule. A regulator can enforce such a rule. A regulator is an organisation that checks whether rules are followed. The source contrasts this with the European AI Act. That law is legally binding. For employees in the Netherlands, the agreement does not change anything straight away. Do not confuse internal agreements with legal duties. Ask your employer which agreements apply to your AI tools. And ask who oversees them.
Background for subscribers
US appeals court rejects appeal on AI training
A US court of appeals issued a ruling. It is called the Third Circuit. The case was Thomson Reuters v. Ross Intelligence. It concerned an AI search engine trained on protected headnotes from Westlaw. The court said this does not count as fair use. Fair use is a US rule. It can allow certain uses of protected work without permission. The court stressed that the case was not about generative AI. That is AI that creates new text or images. Still, lawyers expect other ongoing cases to cite this reasoning. That includes cases about music. This is the first appeals court ruling on this issue. Do you use AI to create text or images? Then no Dutch rule changes straight away. The ruling concerns a US case, a particular product and specific protected material. So do not treat it as general permission or a general ban. But the reasoning may have an impact if parties use it in other cases. Do you work with protected material? Keep track of where it came from. And ask in advance which sources were used.
Background for subscribersAI fake image of political opponent leads to legal threat
In the United States, a member of Congress posted a fake image made with AI. The image falsely attributed views to their opponent. The member of Congress then received a formal demand letter. This is a formal request to do something or stop doing something. This is one of several cases where complaints have followed AI campaign images. The case shows that an image is not harmless just because a computer made it. A convincing image can make it seem that someone said or did something that is not true. If such an image is shared publicly, that person may threaten legal action. The source does not describe a court ruling or mention a new rule. It also does not say what legal grounds the letter relies on. Nor does it say how the case ends. So do not draw a firm conclusion about Dutch law from it. But the same warning applies at work if you publish something on behalf of an organisation. Check the facts and get permission before publishing an AI image. Ask a colleague to check it too.
Background for subscribersEmployers expect tasks to change above all
Workday published a report. In it, 40 per cent of business leaders expect AI to get more out of their current employees. 28 per cent expect fewer staff. These are two different outcomes. According to the report, demand for basic AI skills fell by 25 per cent. This was after a peak in January. Demand for practical skills rose by 51 per cent. This means building AI tools or automating workflows. The report sees a shift for employees. It is less about knowing how AI works and more about using AI in practical ways. The figures describe expectations and skills in demand. They do not predict what will happen to a particular job. So do not read the percentages as a promise of job security or a threat. The figures also do not say what training each employee needs. Choose a task that comes up often this week. Discuss with your team where AI could save time. Also discuss what checks still need to be done. And ask how your employer is introducing AI and what practical knowledge they expect.
Background for subscribersFinancial decision-makers use AI, but often check it
A survey questioned more than 800 financial decision-makers. Of these, 51.5 per cent said they use AI for tax and financial information. That is more than the share who consult the Dutch Tax and Customs Administration. It is slightly less than the use of Google and other search engines. More than six in ten often or always check AI answers against a second source. This helps spot errors. At the same time, 53.6 per cent do not know about the duty to ensure AI literacy. AI literacy means knowing enough about AI to use it properly. According to the source, this duty is in Article 4 of the European AI Act. Only 6.4 per cent say they have already trained all employees. The figures are about financial decision-makers, not all employees. They also do not prove that your employer is falling short. But they do show a gap between use and training. Check an AI answer against a second source this week. And ask your employer what AI training is available.
Background for subscribersAI and tech, made simple.
This is your AI in klare taal edition.
My inklaretaal
inklaretaal · Amsterdam · © 2026