AI Recap: The week AI text in the EU gets a watermark

OpenAI marks ChatGPT text in the EU with an invisible watermark

On Monday OpenAI announced that text from ChatGPT and Codex in the EU will carry an invisible watermark, on all plans, rolling out over the coming weeks. The method is called textGrain, and it lives in the text itself rather than in a file: the model picks its words so that a detector holding a secret key can find the pattern again, even after copy and paste. Readers see nothing, and according to OpenAI the user is not identified. Through the API the feature is available worldwide, but there it is opt-in and off by default. The reason is the EU AI Act, whose transparency rules have applied since August 2 and oblige providers to mark AI content in a machine-readable way. Anthropic had announced in August that it would watermark Claude text worldwide.

OpenAI names the limits itself, and they matter more for daily work than the technique. Replacing around 10 percent of the words with synonyms pushes the detection rate from about 92 down to 66 percent. Short passages, calculation results and translations are hard to detect, and a missing watermark does not prove human authorship. For now only approved researchers and expert organisations get the detector. For companies this means: a proposal drafted with ChatGPT and then thoroughly edited probably no longer carries the mark. And your own duty to tell customers they are talking to an AI in a chatbot or on the phone is untouched by it.

Claude Haiku 5.5: Anthropic's small model now costs a tenth

On Tuesday Anthropic released Claude Haiku 5.5 and priced it at $0.10 per million input tokens and $0.50 per million output tokens, as long as the prompt stays under 100,000 tokens. Haiku 4.5 cost $1 and $5. Above that threshold it gets pricier, $0.50 and $2.50, and a new tokenizer uses slightly more tokens per task; on average Anthropic therefore puts the saving at around 75 percent. The model runs on AWS, Google Cloud and Microsoft Azure. In Anthropic's own benchmarks it lands far ahead of its predecessor (72.4 instead of 15.7 percent on OSWorld, a computer-use test) and, as expected, behind Sonnet 5.5. Anthropic recommends it for high-volume work: summaries, classification, database queries, live support and the subagent role next to the large models. VentureBeat reads the price as a tie with OpenAI's GPT-6 Luna.

For SMBs this is the item with the most direct effect on the bill. Most steps in a typical workflow are exactly this kind of work: file an email, read a receipt, summarise a request, score a contact. Anyone who has been running such steps on a mid-sized model can move a large share of them to Haiku 5.5 and pay a fraction for operations. The large models stay for the spots where judgement counts.

Google gives AI agents their own email address and a place in the company directory

At its Gemini at Work conference on Thursday, Google Cloud CEO Thomas Kurian introduced the "Gemini agent": a single agent for work that answers questions, handles documents, creates images and writes and runs code. You give it an objective instead of a sequence of steps, it plans the work, spins up subagents when needed and delivers the result, even after the laptop has long been closed. It is reachable in the browser, on the phone, in Google Workspace, Microsoft 365 and Slack. The more interesting addition are the coworker agents: persistent agents with a fixed role that get their own email address, their own storage and an entry in the company directory. You add them to a chat space or mention them in a document comment, and they show up in version history under their own name.

What Google sells here as a product are exactly the rules we described as the minimum in our security guide for AI agents: every agent has its own cryptographically attested identity with minimal permissions, sees only the documents shared with it, and every action sits in the audit trail under its own name rather than an employee's. All traffic passes through an agent firewall that applies one central policy to all of them, for example "agents may not open documents classified as confidential". Besides Gemini, Claude from Anthropic may run as the model, and an automatic router picks the cheapest suitable one per task. Google names no prices; the industry editions for financial services and law firms are in preview. According to Google, nearly 90 percent of Fortune 100 companies use Gemini Enterprise.

GPT-6 for every ChatGPT user, plus buttons, calculators and charts inside the answer

Since Wednesday, Pro, Plus, Business and Enterprise customers of ChatGPT have been getting a new GPT-6 model, with free accounts and the Go tier following a day later. Alongside the model OpenAI is rolling out the "Intelligent UI": answers now frequently contain tappable elements, a calculator built for the question, interactive charts or editable graphs. Product manager Aarush Selvan demonstrated a sketch of how an aircraft wing generates lift at launch, plus recipes, a hiking map and a savings calculator. Anyone who dislikes the visuals can dial them back. OpenAI states the goal as making complex topics easier to learn.

For context: GPT-6 Astra started in September in phases, with priority for security firms. Now the GPT-6 generation is open to the whole audience. For companies the news is less the model than the surface: once an assistant's answer is allowed to be a usable tool, the bar moves for internal tools and customer portals too.

Anthropic launches a Cyber Mission with eleven partners and a free scanner for open source

On Thursday Anthropic presented its "Cyber Mission", a long-term commitment to support defenders with models, research and people. Two focus areas at the start: critical infrastructure and open-source software. In the Critical Infrastructure Defense Program, eleven founding partners, among them Accenture, CrowdStrike, Deloitte, Palo Alto Networks, PwC and Rockwell Automation, get access to the frontier models, on-site engineers and threat research to secure operational technology in power grids, water systems and transport networks. For open-source projects there is a free scanner that regularly checks enrolled projects with the strongest models; Anthropic expects a true-positive rate above 90 percent, which is a forecast rather than a measurement. Already on Monday the earlier Project Glasswing, which had scanned hundreds of widely used open-source projects, was folded into the expanded Cyber Verification Program, which gives vetted security professionals access to the most capable models.

Hardly any SMB is directly affected. Indirectly, yes: a large share of the tools automations run on is open source, from n8n to databases to the libraries underneath. If those projects are checked regularly by strong models and findings go privately to the maintainers, holes get closed earlier, before they reach your operations.

What sticks from this week in day-to-day business

First, costs. With Haiku 5.5 the kind of work automations mostly consist of costs a tenth of what it cost a year ago. Anyone running a workflow should go through it once and check which steps really need a large model. Usually it is two or three out of twenty.

Second, rules for agents. Google makes a dedicated identity, minimal permissions and a per-agent audit trail the default and sells it as a product. That is good news, because they are the same rules a self-built agent in n8n needs. If you have not implemented them yet, you will find them in our guide with a checklist.

Third, disclosure. OpenAI's watermark fulfils the provider's duty under Article 50 of the EU AI Act. The duty of the company that deploys a chatbot or an AI phone assistant is a different one: anyone talking to an AI must be able to tell, audibly or visibly at the start of the conversation. No watermark handles that. You do.