AI News
The latest ChatGPT and AI updates, with what changed, who can use it and why it matters for UK readers.
WEEK IN REVIEW · 5–11 NOVEMBER 2026
OPENAI · SMALL BUSINESS · 10 November 2026
OpenAI opens Business Hub in ChatGPT
OpenAI has launched Businesses in ChatGPT in beta. Eligible companies can claim a business profile through Business Hub, verify that they represent it and update the information customers see. OpenAI says verified businesses can also supply product feeds with pricing and availability. Think of it as another place where a customer might encounter your business while comparing their options.
Two parts of the announcement have different eligibility rules. OpenAI says eligible businesses in markets where ChatGPT is available can claim a profile. The interactive Business Agent launches first for eligible ecommerce and online service companies in the United States. A UK shop can check whether it can claim its profile now. The agent’s US launch gives UK owners no confirmed date for when they can add one.
Business Hub reports aggregate data such as mentions, clicks and common discovery topics when activity meets OpenAI’s reporting thresholds. A small shop may therefore have little to measure at first. If you can claim a profile, use your own website as the reference for details you add. Correct anything that could create a bad customer experience, such as an old product price or a delivery area you no longer serve.
For shoppers, the distinction matters too. An agent conversation uses information supplied by the business and is labelled separately from ChatGPT’s independent answers. If you are comparing two suppliers, check the terms and total cost on their own websites before paying. For business owners, the first useful test is to ask a customer-style question and see whether the answer points to accurate information.
Read the original announcement from OpenAI
GOOGLE · AI MODELS · 10 November 2026
Gemini 4 Carbon begins its public rollout
Google has moved Gemini 4 Carbon from its initial announcement into a broader rollout. The company lists access through the Gemini API and its Antigravity development environment, along with a rollout in the Gemini app for Google AI Ultra users. Check your account and plan before assuming you can select it. App access and developer access follow different rules.
One of the changes attracting attention is the ability to delegate work across multiple agents. A useful way to think about this is a project with several independent checks. One part can look for broken links while another checks the wording of error messages. Parallel work can help on tasks like these. It can also make mistakes harder to trace if each agent makes a different assumption about the project.
If you are evaluating Carbon for development, start with a small task you can check independently. Ask for a list of affected files and the tests that support each proposed change. Give the same task to your current model and compare the accepted result, the number of retries and how long you spent reviewing it. The model that finishes first can still take more time overall if its changes need more rework.
Teams using the API should track the total cost of completing that test, including retries and any tools called during the run. The headline price per token is only one part of that calculation. For people using Gemini mainly to draft or summarise text, the more immediate question is whether the features already included in your plan meet your needs.
Read the original announcement from Google
GOOGLE · ANDROID · 10 November 2026
Gemini task automation expands to the UK
Google says Gemini Intelligence task automation is expanding to the UK, India and Japan. The feature can help handle jobs across supported apps, including arranging a ride, ordering food or making a reservation. The same announcement expands Gemini Intelligence support to more phones, including Pixel 10 models. Google says feature availability varies by device, country and language.
A separate feature called Create My Widget is rolling out over the following weeks. It lets you describe a home-screen widget in ordinary language and can draw on information from Gmail, Calendar and Search. A simple starting point would be a widget for upcoming appointments. Check what information appears on the screen if you use the phone at work or share it with someone else.
For UK readers, the announcement is a reason to look for the features on your current phone before considering an upgrade. Two people with similar devices can see different options as a rollout progresses. Check the exact model, the app you want to use and Google’s current availability information if the feature is missing.
If you try a reservation or order, compare the details Gemini has selected with what you originally asked for. Watch for date, time, address and total price. A good first trial is a simple request you can review in the original app. With a widget, look at whether it saves you a step you repeat every day. The useful test is how it fits your routine once the novelty wears off.
Read the original announcement from Google
ANTHROPIC · AI MODELS · 9 November 2026
Anthropic releases Claude Fable 5.2 and Mythos 5.2
Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. It describes them as versions of the same underlying model with different safeguards and access arrangements. Fable has additional protections for broader use. Mythos is restricted to vetted users. The similar names are worth noticing when you read a benchmark, a review or a developer’s account of using one of them.
Anthropic’s announcement covers coding, scientific and engineering work, and cybersecurity. For many readers, the practical question is whether Fable helps with a task that regularly gets stuck. For example, you might ask it to explain why two reports disagree, or to trace a bug across a small codebase. Give it the same starting material you would give your existing model so you can compare the answers.
Decide what a successful result would look like before you run that comparison. In a code review, you might ask the model to point to the relevant lines and explain how to reproduce the issue. In a document review, ask it to show the paragraphs that support its conclusion. A polished explanation is easy to read, but the evidence is what lets you accept it or send it back for revision.
Check which model you can actually select in the Claude product or API you use. For workplace tests, start with example material you are allowed to share and check your organisation’s rules before uploading real files. If a simpler model already gives you accurate, easy-to-check answers, keep the new release in reserve for the jobs where the extra review effort is worth it.
Read the original announcement from Anthropic
OPENAI · CYBERSECURITY · 5 November 2026
OpenAI introduces GPT‑6 Cyber for authorised research
OpenAI introduced GPT‑6 Cyber for authorised security research. It is designed to help qualified teams investigate software vulnerabilities, check whether a finding can be reproduced in an approved test environment and verify candidate fixes. OpenAI provides restricted access through Daybreak Red. The release is aimed at security practitioners working within agreed boundaries.
OpenAI reports an ExploitGym score of 48.5% for GPT‑6 Cyber, compared with 32.1% for GPT‑5.6 Cyber on the same evaluation. That is a 16.4 percentage-point difference. ExploitGym measures performance on specified test vulnerabilities. Results on that benchmark tell us how the models compared under those conditions. A team would still need to assess coverage and false alarms in its own environment.
OpenAI also describes using the model in research involving Firefox and says Mozilla has released fixes for many of the reported vulnerabilities. For most readers, the part of that story that changes what to do today is simple. Keep browsers and other regularly used software up to date so fixes reach your device. Software vendors’ release notes can help you see when a relevant patch is available.
For security leads assessing tools like this, ask how the system records the scope of an assessment and who can review its actions. Agree in advance on which systems can be tested, how to stop a run and where findings should go. A good outcome is a report that someone can reproduce, prioritise and use to verify a fix. The model’s benchmark score is one part of that decision.