AI news for September 2026: what changes for businesses

September 2026 was a busy month. The AI news touched the rules, both European and Italian, and then the models and their prices. In this article I collect the AI news that, in my view, matters most to a business. For each item I give the official source and the reason it deserves attention.

One caveat, though: I report facts published by 26 September 2026. Prices and dates are those stated by vendors and institutions at that time.

AI news from Europe: transparency rules now apply

Since 2 August 2026, Article 50 of the AI Act applies. It sets the transparency duties. The European Commission has also published its final guidelines; the page was last updated on 6 August 2026.

For organisations that use AI systems, which means many companies, there are therefore three main duties. First, tell people when they are exposed to emotion recognition or biometric categorisation. Next, disclose deepfakes. Finally, flag AI-generated text on matters of public interest when it is published without human review or editorial control. Providers of the systems, instead, must tell people when they are talking to an AI and mark generated content in a machine-readable way.

Vendors, in the meantime, have moved too. On 14 August Anthropic explained how the watermark in Claude's text works, added to comply with the AI Act. According to Anthropic, it adds no hidden characters and cannot be traced to a specific person.

Meanwhile, in other EU AI news, the timeline for high-risk systems has changed. On 7 May 2026 the Council and the European Parliament agreed on new dates: 2 December 2027 for stand-alone systems and 2 August 2028 for systems embedded in products. The deal also shortened the grace period for marking generated content; the new deadline for providers is 2 December 2026. The final text is Regulation (EU) 2026/1744, in force since 27 July 2026 according to the law firm Orrick.

Why it matters. A chat assistant on your site, an AI image for a campaign, a press release drafted with a model: these are everyday cases. So it pays to list your uses now and decide, for each one, what must be disclosed.

Italy: an implementing decree is published

Among the Italian AI news, on 15 September 2026 the Official Gazette no. 214 published Legislative Decree no. 160 of 9 September 2026. It covers AI in police work and civil and criminal liability. It enters into force on 30 September 2026.

The Council of Ministers statement of 4 August 2026 sums it up. For businesses, two points stand out. First, a new Article 437-bis of the Criminal Code punishes failing to adopt security measures in high-risk AI systems. Second, people who suffer harm get stronger protection: access to the system's technical documentation, a presumed causal link and a direct claim against the insurer.

At the same meeting the government gave final approval to a second decree, on national authorities and training. According to the statement, the Agency for Digital Italy will be the notifying authority and the National Cybersecurity Agency the market surveillance authority. Moreover, decisions on hiring, discipline and dismissal may not rest solely on automated processing. I have not yet verified whether this second decree has been published in the Gazette.

Why it matters. In other words, if you provide or use high-risk systems, you should document your security measures now. And if you use AI in HR, the final decision will have to stay with a person.

AI news from vendors: new models, lower prices

OpenAI. On 3 September it launched GPT-6 Astra, its flagship model. Its API price list sets $10 per million input tokens and $50 per million output tokens. On 22 September came GPT-6 Sol and GPT-6 Luna, faster and cheaper. OpenAI says prices are half the GPT-5.6 promotional rates. Per million tokens, Sol costs $2 in and $10 out, and Luna $0.10 and $0.50.

Anthropic. On 1 September it announced Enterprise Frontier Safeguards, which combine the privacy of zero data retention with misuse detection. It will roll out in phases, with the goal of broad availability later this fall. Also in September it released Claude Fable 5.1. Thanks to lower cache read prices, Anthropic estimates about 25% lower costs for typical workloads. On 22 September came Claude Opus 5.5: $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5.

Google. According to the Gemini API release notes, Gemini 3.8 Flash became generally available on 2 September. On 17 September, in addition, Google released a new version of its Antigravity agent with renamed tools; the old version shuts down on 5 October. Then, on 18 September, it limited access to Gemini 2.5 models to users who had used them before.

Why it matters. The AI news from vendors cuts both ways: prices fall, but the models under your projects change too. An integration built for one model may need to migrate within weeks. As a result, you need pinned model versions, repeatable tests on real cases and a steady eye on retirement dates.

AI news in practice: what I would do this month

If I ran a company that uses AI, in light of this AI news I would do five things:

1. list every AI use, who uses it and for what;
2. for each use, check the Article 50 transparency duties;
3. review HR decisions, so that none rests only on an automated system;
4. recalculate the costs of active projects against the new price lists;
5. list the models in use, with the retirement dates vendors have announced.

These are short tasks, yet they prevent surprises. They also give you a solid basis for deciding where to invest in the coming months.

On the page about AI-assisted data processing I describe how I use AI to process data: bounded use cases, clear test criteria and transparent costs.

If you want to know which of these changes really affect your company, write to me and describe how you use AI today. I will reply with a concrete reading, point by point; if a broader piece of work is needed, the page on technical consulting on data and tools explains how I set it up.