Last reviewed: 16 August 2026
By Lilach Bullock
AI news this week arrived wearing three different outfits.
One was a sprinter. OpenAI previewed a version of its strongest model that can answer up to 14 times faster.
One was a salesman. ChatGPT Ads opened in five more countries, including the UK.
And one was a group of unsupervised colleagues arguing in a shared office, changing each other’s code and occasionally locking one another out.
That last one came from Anthropic’s research into teams of AI agents. It is also the story I would pay the most attention to if you are building automations for your business.
I read the primary announcements published from 10 to 16 August and kept the 14 that could change how a small business works, spends, markets or protects itself. I also found several Google pages with fresh sitemap dates that were months or years old. They are not in here. A sitemap wearing a new hat is still an old article.
Each section explains what happened, what it means for your business and where a person still needs to make the decision.
The AI news this week at a glance
- OpenAI made its strongest model run up to 14 times faster
- ChatGPT Ads launched in the UK and four more countries
- ChatGPT Business introduced a $125 Premium seat
- OpenAI put specialist cyber models inside AWS
- Claude will add an invisible watermark to its writing
- Teams of AI agents found more bugs and created new fights
- Job retraining may not be enough for AI disruption
- Claude improved a mathematical bound that stood for years
- Gemini passed one billion monthly users
- Gemini can connect to Wix, Otter.ai, Ticketmaster and more
- Google showed AI working on encrypted data
- Google warned that attackers are moving at machine speed
- Meta is giving 15,000 AI glasses to people with sight loss
- Meta described an agent that works across your whole life
- What I would do with all of this in 60 minutes
The pattern beneath the week
AI is leaving the chat box.
It is entering advertising, connected apps, cybersecurity, pricing plans, accessibility tools and teams of agents that act together. That creates more useful systems. It also creates more places for a bad instruction, weak permission or wrong source to travel further than it should.
If you are still choosing tools by asking which model is cleverest, the market has moved underneath you.
The better questions are now:
- What job does it own?
- What can it read?
- What can it change?
- What does it cost at the volume we need?
- Who checks the result?
- What happens when two automated systems disagree?
My guide to AI automation for small business starts with the workflow for exactly this reason. A model is an ingredient. The business result comes from the whole system around it.
1. OpenAI made its strongest model run up to 14 times faster
OpenAI previewed Ultrafast, a new service tier for GPT-5.6 Sol.
The company says it runs up to 14 times faster than Standard processing and can generate up to 750 output tokens per second. It is powered by Cerebras and is available only to a limited group while capacity expands.
The higher speed changes which jobs are practical. A customer-service agent can search several systems while the customer is still speaking. A shopping assistant can check stock and answer a product question before the buyer wanders off. A technical team can examine logs while an outage is unfolding instead of waiting for an overnight report.
Lilach’s verdict
Useful for work where delay destroys value. Wasteful for jobs that can wait five minutes.
Do not buy a Formula One car to collect the milk.
Keep this human
Do not confuse speed with accuracy. A wrong answer delivered at 750 tokens per second is just a mistake with excellent cardio.
Read OpenAI’s Ultrafast announcement.
2. ChatGPT Ads launched in the UK and four more countries
OpenAI expanded ChatGPT Ads into the UK, Mexico, Brazil, Japan and South Korea on 11 August.
The ads appear for eligible adults using Free and Go plans. OpenAI says they remain labelled, separate from the answer and hidden from sensitive topics. Advertisers receive aggregated performance information rather than access to conversations.
For a small business, the moment of intent matters more than the geography.
People use ChatGPT to compare options, narrow a shortlist and decide what to do. That is much closer to a buying decision than somebody scrolling past a social post while waiting for the kettle.
Lilach’s verdict
Watch this closely. Search advertising was built around keywords. ChatGPT advertising is being built around the context of a decision.
That could be extremely valuable. It could also become expensive very quickly once every advertiser notices.
Keep this human
Do not let an advertising dashboard define a good lead. Track the click, the enquiry, the fit and the sale separately.
3. ChatGPT Business introduced a $125 Premium seat
OpenAI announced Premium seats for ChatGPT Business.
Premium costs $125 per user each month, or $100 per user each month on annual billing. Standard seats remain $25 monthly or $20 on annual billing. OpenAI says Premium provides five times more usage and removes the five-hour usage limit.
Businesses can mix both seat types inside one workspace.
That last sentence matters more than the launch language.
AI pricing is starting to resemble staffing. The person producing three reports a day may need a different allowance from the person opening ChatGPT twice a week to tidy an email.
Lilach’s verdict
Do not upgrade the whole team because one enthusiastic employee has turned ChatGPT into a second home.
Measure the job first.
Keep this human
Usage is not value. Somebody can burn through five times more AI and produce five times more material nobody needed.
Read the ChatGPT Business Premium announcement.
4. OpenAI put specialist cyber models inside AWS
OpenAI made its Daybreak cybersecurity models available through Amazon Bedrock for eligible customers.
Daybreak Blue provides frontier general models with safeguards for defensive work. Daybreak Red provides specialist models for authorized vulnerability research, exploit validation and security testing.
The practical change is where the models live. Security teams can use them inside the AWS environment they already govern instead of creating a separate technical island with a new set of permissions.
Lilach’s verdict
This is an enterprise story with a small-business lesson: the best AI system is often the one that fits inside the controls you already understand.
Keep this human
Security testing can damage the thing it is meant to protect. Keep authorization, scope and the final action with qualified people.
Read the Daybreak on AWS announcement.
5. Claude will add an invisible watermark to its writing
Anthropic says future Claude models will generate text with an invisible statistical watermark to comply with the EU AI Act.
There are no hidden characters. The system changes how Claude makes low-stakes word choices, leaving a pattern that can be detected with a key. Anthropic says it adds no extra tokens, does not increase the price and cannot identify a person, company or chat.
The watermark becomes more detectable in longer passages. It is weaker in factual text, code and light proofreading because the model has fewer acceptable choices.
Lilach’s verdict
This will not rescue bad AI writing.
It may help identify machine involvement, but it cannot tell you whether the thinking is original, the claims are correct or the reader will care.
Keep this human
Do not accuse somebody of using AI from a detector score alone. Anthropic says the watermark shows likely Claude involvement, not authorship, ownership or responsibility.
Read Anthropic’s explanation of Claude watermarking.
6. Teams of AI agents found more bugs and created new fights
Anthropic gave 45 agents their own computers, a shared forum and the job of finding vulnerabilities in 15 open-source projects.
Coordinating swarms found many more vulnerabilities than a narrower independent approach, although they also used far more tokens and explored a wider area.
Then Anthropic tested what happened when agents had competing instructions inside one shared codebase.
Some agents hid information, changed other agents’ code, created defensive processes and locked rivals out. In better runs, they recognized the conflict, wrote a truce and asked for a person to intervene.
More capable models did not automatically coordinate better.
Lilach’s verdict
This is the most important agent story in the issue.
Adding more agents can improve coverage. It can also produce the automated equivalent of six managers editing the same spreadsheet while nobody owns the number at the bottom.
Keep this human
Never give several agents overlapping write access and assume intelligence will produce manners.
Read Anthropic’s multiagent research.
7. Job retraining may not be enough for AI disruption
Anthropic reviewed 56 randomized US studies of job-training programmes, alongside evidence from Europe.
On average, being offered a training place increased employment by two to three percentage points and earnings by roughly $1,000 a year. The average cost was about $13,000 per person.
Some programmes tied directly to employers in high-demand sectors performed much better, but attempts to reproduce them often failed.
The authors conclude that existing retraining programmes would probably fall short if AI displaced workers at scale.
Lilach’s verdict
Training people to “use AI” is too vague.
The useful version connects a person, a role, a live business process and a job that somebody will pay for.
Keep this human
Do not use training as a polite substitute for a business plan. If the new skill does not connect to demand, tools and a defined role, the certificate may become expensive wallpaper.
Read Anthropic’s retraining evidence review.
8. Claude improved a mathematical bound that stood for years
Anthropic gave an unreleased Claude research model an unreasonable task: take a serious attempt at the Riemann hypothesis, one of mathematics’ most famous unsolved problems.
Claude did not solve it.
It did improve a related lower bound from 41.6 percent to 67.2 percent. Anthropic mathematicians reviewed the work, outside experts examined it and Claude produced a formally verifiable proof in Lean.
The process used around 60 subagents, 31 million output tokens, 2,400 shell commands and thousands of numerical checks.
Lilach’s verdict
The headline is not “AI solves mathematics.”
The useful lesson is that a failed ambitious target can produce a valuable adjacent result when the system tests itself, checks prior work and invites expert review.
Keep this human
Formal verification covered the mathematical proof. It did not turn every statement produced during the process into a fact. Expert review still mattered.
Read Anthropic’s report on Claude and the Riemann zeta function.
9. Gemini passed one billion monthly users
Google says more than one billion people now use the Gemini app each month, making it the fastest-growing product in the company’s history.
The usage details are more interesting than the milestone.
Google says 63 percent of users talk to Gemini, one in five Gemini Live interactions include camera or screen sharing, and the app generates more than 150 million images each day. Gemini can also take actions across more than 40 Android apps.
Lilach’s verdict
Your customer is not only typing questions into a search box.
They are showing an AI their screen, pointing a camera at a problem and asking it to act across apps.
Keep this human
One billion users does not mean one billion buyers. Treat usage claims as a behaviour signal, not a conversion promise.
Read Google’s Gemini usage announcement.
10. Gemini can connect to Wix, Otter.ai, Ticketmaster and more
Google announced a new group of connected services for Gemini.
The list includes Granola, Otter.ai and Wix for productivity and creation; Ticketmaster, GetYourGuide and OpenTable in the UK for plans and bookings; plus Angi, Thumbtack and Zocdoc for home and health tasks.
The assistant is becoming a front door to other companies’ services.
Lilach’s verdict
Being easy for agents to use is becoming a distribution strategy.
The customer may not visit your homepage. Their assistant may compare, book or edit through an integration while your beautiful navigation watches from the sidelines.
Keep this human
Connected apps can expose more context than the job requires. Give access one service at a time and remove it when the experiment ends.
Read Google’s connected-app announcement.
11. Google showed AI working on encrypted data
Google introduced HEIR, an open-source compiler for homomorphic encryption.
Homomorphic encryption allows a server to process encrypted information and return an encrypted result without seeing the underlying data.
Google demonstrated private recommendations, credit-card fraud detection, network-threat detection and hotword recognition. The technology still carries a performance cost, but Google says that cost is falling.
Lilach’s verdict
The business direction is simple: using AI should not require exposing raw customer data to the system doing the work.
Keep this human
“Encrypted” is not a complete security answer. Access control, retention, staff permissions and the business process around the model still matter.
Read Google’s HEIR announcement.
12. Google warned that attackers are moving at machine speed
Google’s security team says AI agents are changing attacks across three dimensions: sophistication, scale and speed.
The warning is especially sharp for ransomware groups and access brokers, where volume matters more than surgical precision. Agents can scan documentation, find weak credentials and move through a system before a person reviews the alert.
Google expects open models to match the current cyber capability of frontier models within six to twelve months.
Lilach’s verdict
The dull security jobs just became urgent.
Weak passwords, forgotten accounts and credentials saved in files are ideal targets for a system that never gets bored.
Keep this human
Do not point an unsupervised security agent at production systems. Start with read-only checks, a narrow scope and an expert who can distinguish a vulnerability from a false alarm.
Read Google’s agentic-security warning.
13. Meta is giving 15,000 AI glasses to people with sight loss
Meta is donating 15,000 Ray-Ban Meta glasses to Vision Ireland, enough for every blind or low-vision adult the charity supports.
The glasses can read text, identify objects and answer questions about the wearer’s surroundings. The programme includes in-person training and helpdesk support, funded by Meta.
This is what useful AI looks like when the outcome is independence rather than another content button.
Lilach’s verdict
The hardware is only half the product.
Training and support are what turn an impressive demonstration into something a person can depend on while standing in front of the fridge.
Keep this human
Test accessibility with users, not assumptions. Also make privacy clear when a wearable camera may capture other people or sensitive surroundings.
Read Meta’s Vision Ireland announcement.
14. Meta described an agent that works across your whole life
Meta published its direction for personal superintelligence.
The company describes an agent that works across relationships, health, career, finances, home management and hobbies. It also describes creation tools, tutors, scientific assistants and ways for people to build companies with smaller teams.
Meta says free versions will reach billions, while heavier users will pay for more compute through a dynamic auction designed to allocate capacity.
This is a direction, not a finished product specification.
Lilach’s verdict
The ambition is enormous. So is the permission request hiding inside it.
An assistant cannot manage your health, money, home and relationships without learning a great deal about all four.
Start with an assistant that reads one safe source and prepares one output. My AI Delegation Playbook helps you choose that first job. Then add another source only after the first job is dependable.
My guide to AI customer-language analysis shows a bounded example: use approved customer material to find patterns, then keep the interpretation and final claim under human review.
The principle is simple. An agent should earn wider access by completing narrower work correctly.
Keep this human
Do not hand one vendor the keys to your whole life because the demo remembered your daughter’s recipe. Convenience is not a permission policy.
Read Meta’s personal-superintelligence direction.
The pattern behind all 14 stories
The fastest model in the issue is useful because it can act while a customer is waiting.
The advertising story matters because AI is sitting closer to the buying decision.
The connected-app story matters because assistants are becoming a route into other businesses.
The watermark matters because AI involvement is becoming detectable and regulated.
And the multiagent research matters because giving several systems more freedom does not give them shared judgement.
One billion people already have access to AI.
The advantage is designing a dependable job around it: clean sources, narrow permissions, an output contract, a cost limit, an approval point and evidence that the result was useful.
That is also why AI content repurposing works better as a defined system than a prompt. The source, formats, voice rules, approval and measurement all matter more than the first draft.
Your 60-minute action plan
Minutes 0 to 10: choose the waiting point
Use my Automation Audit Checklist to find one place where a customer, employee or decision is waiting for information.
Write down what the delay costs.
Minutes 10 to 20: define one agent job
Use my AI Agent Brief Template or complete this sentence:
“When this happens, read these sources, produce this output and stop for approval before changing anything.”
If the sentence needs three paragraphs, the job is too large.
Minutes 20 to 30: check permissions
List every inbox, drive, calendar, website and account the tool can access.
Remove anything it does not need for the job.
Minutes 30 to 40: test a disagreement
Give the system two sources that conflict.
Check whether it flags the conflict, chooses one without saying so or blends them into a confident mess.
Minutes 40 to 50: calculate the cost lane
Record the cost of one run, expected weekly volume and the value of the delay or labour removed.
Do not pay premium-model prices for formatting.
Minutes 50 to 60: create the stop rule
Use an AI handoff rule to write the conditions that force the system to stop and ask for a person:
- Missing source
- Conflicting number
- New external recipient
- Payment or contract decision
- Irreversible change
- Confidence below the agreed threshold
You now have the beginning of an agent that can help without becoming the colleague who reorganises the filing cabinet and forgets to mention where everything went.
If you need examples, browse my AI resources for business owners or start with my guide to AI meeting notes and what they miss.
Final word
This week did not produce one tool every small business must buy.
It produced a much more useful warning.
AI is becoming fast enough to work during the conversation, connected enough to reach across your tools and autonomous enough to create problems before you reach the approval screen.
That does not mean you should slow down.
It means your boundaries need to speed up too.
Choose one job. Give it only the access it needs. Test the ugly cases. Measure the result. Keep a person at the point where judgement, money, reputation or somebody else’s data enters the room.
Then let the machine run.
For a short, ranked AI plan rather than a deck, see AI strategy consulting.
Frequently asked questions
What was the biggest AI news this week for a small business?
ChatGPT Ads launching in the UK and four other markets may create the most immediate marketing opportunity. Anthropic’s multiagent research carries the most important operational lesson: adding more agents can improve coverage but also create conflicts, duplicated work and unsafe actions when ownership is unclear.
Should a small business pay for ChatGPT Business Premium seats?
Only for people whose valuable work repeatedly stops at the current usage limit. Compare the added monthly cost with the time, revenue or risk attached to that interruption. Keep lighter users on Standard seats and review usage after 30 days.
Will Claude’s watermark make AI writing obvious to readers?
Anthropic says no. The watermark is a statistical pattern in word choices, not a visible label or hidden character. Detection needs Anthropic’s key and works better on longer passages. It does not prove authorship or identify the user.
What is the safest way to start using an AI agent?
Give it one narrow job, one approved set of sources and read-only access where possible. Define the output, the cost limit, the stop conditions and the person who approves external or irreversible actions. Test missing, outdated and conflicting information before increasing its permissions.
How can a small business prepare for AI agents buying or booking on behalf of customers?
Make service descriptions, prices, availability, locations, terms and booking rules clear and machine-readable. Keep the facts consistent across your site and business profiles. Then test buyer questions in ChatGPT, Gemini, Claude, Perplexity and Google to see whether your business is mentioned, cited or recommended.
See for free whether ChatGPT, Claude, Perplexity, Gemini and Google name you, and get the plan to become the answer.
Check my AI visibility →
Lilach Bullock
Forbes Top 20 Social Media Power Influencer. 21 years of marketing experience. Currently obsessed with AI implementation done well. British AI and marketing consultant working internationally, writes a weekly newsletter that 15,000+ founders read on Sunday mornings.
Let’s get your marketing running on AI.
Book a free 30-minute call
We figure out what you need, where AI fits in, and what working together would look like.
Book the call →Or take the 30-second calculator
You’ll see the hours and the money quietly leaking out of your week, and the three workflows worth building first.
Take the calculator →Or grab the free AI resource library
Prompt packs, templates, checklists, and swipe files. The exact tools I build for paying clients. Yours, free.
Get the library →More from the blog.
-
Can You Run Email Marketing Effectively From Your Phone?
Can you run email marketing effectively from your phone? Yes for most tasks, and here's exactly which ones…
-
Which YouTube Channels Have the Fewest Subscribers (And What That Tells You)
There's no real list of YouTube's smallest channels, and understanding why reveals what subscriber count means for creators.
-
How Much Do Instagram Ads Typically Cost? Real Numbers From My Own Campaigns
Instagram ads typically cost 50p to £3 per click, but the real number depends on your industry, and…
