Anthropic and OpenAI cut prices as access to AI gets contested
The price of top-tier AI models became easy to judge this week: the two leading labs cut it on the same day, with the numbers on their websites. Access became harder to judge. Who may test a model before release, and whose agent may act on a website, is now decided by other companies and by governments, with little notice to the businesses that depend on them.
Anthropic and OpenAI cut prices on their top AI models the same day
On September 22 Anthropic released Claude Opus 5.5 at token prices 20% below Opus 5, and said it performs at the level of Claude Fable 5.1, its most capable model, on most work. The same day OpenAI released GPT-6 Sol at half the per-token price of the GPT-5.6 model it replaces, plus a cheaper GPT-6 Luna for high-volume tasks.
Both cuts land on the tier most companies buy for demanding work. With list prices this close, what separates the two is cost per finished task: a cheaper model that needs a second attempt ends up costing more.
Source: InfoWorld
Amazon blocks Meta's Muse shopping agent from Amazon.com
On September 20 Amazon began stopping Muse, Meta's consumer AI assistant that buys things on a user's behalf, from completing purchases on its store, less than two weeks after Muse launched in the US. Amazon said it acted after Meta declined to withdraw the agent, which does not identify itself when it browses.
An agent reaches only the stores that agree to let it in, and those agreements are now being made one retailer at a time. A brand that sells through marketplaces should know which agents can reach its listings, and on whose terms.
Source: Bloomberg
The White House asks OpenAI and Anthropic to delay UK model testing
The US Office of the National Cyber Director asked OpenAI and Anthropic to hold new models back from the UK AI Security Institute, Britain's government testing body, until a US security review is done. Politico reported the request on September 24, and a British official confirmed it to Bloomberg.
For companies in Europe, independent checks on new US models may now come later, after a US review.
Source: The Next Web
Finland and Norway lead a call to test powerful AI before release
Finland and Norway launched "A Call for Control of Frontier AI Models" on September 21. More than twenty countries have signed it, Germany and Canada among them. The European Union signed as well. It asks for mandatory testing before release, independent evaluators with real access to the models, shared reporting of incidents, and possibly an international oversight body.
The US and China, which build most leading models, stayed out, and so did the UK. For companies in Europe little changes: the EU AI Act already requires the makers of the most capable models to test them, including adversarial testing, and to report to the EU's AI Office.
Source: Government of the Netherlands
One in five UK firms cut entry-level hiring, an Open University survey finds
In a survey of 1,500 UK business leaders published in September, the Open University found that 19% have reduced entry-level recruitment, and 42% of those name AI adoption as a reason.
The routine tasks that used to train juniors are the first ones handed to AI, which leaves firms with fewer ways to train their future managers.
Source: CNBC
Brazil's Supreme Court catches a lawyer's hidden AI command in a filing
On September 25 Justice Cristiano Zanin of Brazil's Supreme Federal Court ordered an investigation into a filing that carried a hidden instruction written for AI, meant to make the court's electronic system accept an appeal automatically. A tool built by the court's own IT team caught it. The Brazilian Bar Association will consider disciplining the lawyer.
This is prompt injection: text planted in a document to redirect the AI that reads it. It came from a lawyer in the case itself. Any firm that uses AI to sort incoming documents should test for it.
Source: Agência Brasil
Before the next price cut, write down which models your teams use, who tested them before release, and which platforms your agents need to reach. Price has become the easy part of the choice. The next surprise is more likely to come from access: who checked the model, and which sites let it act, both decided by parties you have no contract with.