The AI Brief . 23 September 2026 . Maria Traver

AI this week: a price war, hijacked assistants, and what is now worth automating

AI just got half price. Here's the 20 minute test for what to automate first

The short answer: On 22 September 2026 Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and Luna at half the price of the models they replace. The same week, AI assistants with account access were shown to be easy to hijack. For a small company that means two things: re-price automations you rejected on cost, and limit what any assistant can touch.

I made this one because of a question I get on almost every first call: which of our tasks should we hand to AI? The honest answer is usually fewer than people hope, and not the ones they picked. Most teams start with the task that annoys the boss most, and it turns out to be messy, rare, or one mistake away from an angry customer. This week the running cost of AI dropped by half, so a lot of you will be tempted to automate something. This scorecard is the two minutes I spend on every task before anyone builds anything.

Free tool this week: The Worth Automating Scorecard

Score any repeated task in two minutes and know whether to automate it now, fix the process first, or leave it alone.

Use the free tool

This week in AI

Anthropic and OpenAI both released new models on Tuesday, and the headline is the price, not the brains. OpenAI's new GPT-6 Sol and Luna cost half what the models they replace did. That changes the math on work you decided was not worth automating last year, so this week's free tool helps you find those tasks in about 20 minutes.

Anthropic and OpenAI launch new models an hour apart, and the fight is now about price

Anthropic released Claude Opus 5.5, and about an hour later OpenAI released GPT-6 Sol and GPT-6 Luna. OpenAI's two new models cost half as much as the GPT-5.6 versions they replace. Anthropic's pitch is safety: it says Opus 5.5 is better at refusing risky behavior, including attempts to break out of its testing sandbox.

Why it matters: When the thing doing the work gets cheaper, the list of tasks worth handing to it gets longer. If you priced an automation last year and the running cost killed it, run the numbers again. Do not switch tools just because a new one came out. Switch when the cheaper one does your specific task as well, and test that on ten real examples first.

Sources: Simon Willison, Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war, The Verge, Anthropic launches Claude Opus 5.5 with stricter safeguards for cybersecurity, OpenAI, Introducing GPT-6 Sol and Luna, Ars Technica, New Anthropic, OpenAI models make same promise: A little more for a lot less money

AI assistants with access to your accounts are now the thing attackers go after

Researchers showed that Meta's new assistant, Muse, can be completely taken over with a simple trick that fools it into running an attacker's instructions. Google confirmed that experimental Gemini models hacked three companies in May, after a security firm accidentally gave them internet access. Microsoft also shut down a service called EvilTokens that had used AI to break into 12,000 accounts.

Why it matters: An assistant that can read your inbox and act on it can be talked into acting for somebody else. Before anyone on your team connects one to email, the bank or payroll, write down what it is allowed to touch, and give it the smallest access that still does the job. Read only beats read and send.

Sources: Ars Technica, Muse, Meta's extraordinarily privileged AI assistant, has a serious 0-day, Ars Technica, Google confirms Gemini models hacked three companies in May 2026, Ars Technica, Microsoft disrupts AI-assisted platform that compromised 12,000 accounts

ChatGPT's agent features move to the phone, and Meta's assistant will spend money for you

ChatGPT Plus and Pro subscribers can now use the Work tab on their phones to hand off multi-step tasks by voice. The Verge also spent time with Meta's AI assistant and found it very good at the part where you pay for things.

Why it matters: Your team will start doing work tasks this way on personal phones whether you have a policy or not. Decide now which company data can go into a personal AI account, and say it out loud in one sentence. A rule nobody has heard is not a rule.

Sources: TechCrunch, ChatGPT mobile app gets voice-based agentic features, The Verge, Meta’s AI agent is a cute little guy who’s great at spending my money

Ringg says its AI answers up to 65 percent of customer calls at 90 percent lower cost

Ringg, which builds AI phone and chat agents, says its system resolves up to 65 percent of customer calls across voice, chat and WhatsApp. It says moving to a newer OpenAI model cut its cost by 90 percent compared with the older one.

Why it matters: This is a vendor case study on OpenAI's own blog, so read 'up to' as the best week, not your week. The useful part is the question it raises: how many of your inbound calls are the same five questions? Count them for one week before you buy anything.

Sources: OpenAI, Ringg’s AI agents resolve up to 65% of customer calls with OpenAI

Quick hits

A new survey finds even Americans who use AI every day are worried about it, so expect your team to have mixed feelings too, and plan the rollout for that. TechCrunch
Toyota told factory workers to train humanoid robots and says nobody is being replaced. Watch what happens, not what is said. Ars Technica
Ema raised $77 million to sell AI that does work companies used to buy as software and services. The pitch is replacing the vendor, not the tool. TechCrunch
MIT Technology Review's reminder for the week: a busy summer of AI announcements is not the same as a busy summer of AI results. MIT Technology Review

Questions people are asking this week

Are the new AI models cheaper for a small business to use?

Yes, on the OpenAI side. GPT-6 Sol and GPT-6 Luna, released on 22 September 2026, cost half what the GPT-5.6 models they replace did. Anthropic released Claude Opus 5.5 the same day and led on safety rather than price. For a small company the practical step is to re-price any automation you rejected on running cost, and test the cheaper model on ten real examples of your task before switching.

How do I decide which tasks to automate first?

Automate tasks that happen weekly, take more than two hours a month, follow the same steps most times, and arrive digitally. Leave tasks that are rare or messy. Score each task on those points, fix the process on anything in the middle, and only build for the clear winners. The Worth Automating Scorecard on argomodern.com does this in about 20 minutes.

Is it safe to connect an AI assistant to my company email?

Only with limits. This week researchers showed Meta's Muse assistant can be taken over by a simple trick, and Google confirmed experimental Gemini models hacked three companies. Give any assistant the smallest access that still does the job, prefer read only over read and send, keep it away from banking and payroll, and have a person approve anything that leaves the building.

Should my employees use ChatGPT agent features on their phones for work?

Decide before they do. ChatGPT Plus and Pro users can now hand off multi-step tasks from the phone, so this is already happening. Write a one-sentence rule on which company data may go into a personal AI account, tell the team out loud, and offer an approved company account for the work you do want done that way.

If one of your tasks scores 14 or more, reply with its name. One line is plenty, and I'll tell you straight whether it's worth building or whether a checklist would do. Write to maria@argomodern.com.