In this issue
A new model named Jev has arrived, developed by a founder who worked on the research behind ChatGPT. You send it your data and your questions. It answers each one with a typed value and a probability. Input costs four cents per million tokens and output is free. This issue carries a full section on it, with the news so far and where to go deeper.
The other half of the issue is security. Three researchers broke into OpenAI using a tool built by Anthropic. Google confirmed that Gemini guessed its way into three real companies during a test. Claude Cowork and chat merged into one app. OpenAI put a chat agent inside its ads. A coffee shop in Buffalo found out what its customers think of AI art. Let's get into it.
Topics of the day:
Jev decides instead of writing, for almost nothing
Researchers broke into OpenAI with Claude
Claude Cowork and chat become one app
ChatGPT ads can now chat with your customers
Curated reads on agent skills, agent teams and Stripe's Kai
ChatGPT's Data agent builds dashboards on request
An AI menu poster cost a coffee shop followers
The Shortlist: OpenAI's incident log, Astra for Law, Grok
Jev decides instead of writing
What's happening: TypeSafe released Jev on September 15, after two years in stealth. Its founder, Diogo Almeida, worked at OpenAI on the research behind ChatGPT.
You send it a state, which is your data as structured program state, plus a list of questions. It returns a typed answer to each question with a probability attached. TypeSafe calls this class a System One model, after Daniel Kahneman's fast thinking. It trained Jev with a method it calls reinforcement learning for calibrated decisions.
The numbers are the reason people are talking. Input costs $0.042 per million tokens, which is $42 per billion, and output tokens are free. TypeSafe says a call finishes in 70 to 500 milliseconds.
In practice:
Point it at the decisions inside your workflows, classify a ticket, route a lead, score a form, pull fields out of a message.
Read the probability on every answer, and send the uncertain ones to a person instead of acting on them.
Keep your chat model for the writing, because Jev returns typed values and never prose.
Budget setup time, you define the questions and the allowed answers before the call, which is more work than writing a prompt.
Bottom line: The hundred small decisions inside your workflow no longer need a chat model.
Read the claims carefully: TypeSafe says Jev cannot return a value outside the answer shape you define. So it cannot invent a field or a type, and that part is checkable. The multipliers on its home page are a different matter. The 193.6 times faster and 444.6 times cheaper come from workflow tests TypeSafe wrote itself. It scored them against the average of GPT-6 Astra and Fable 5.1, and it says so on the page. I would treat those two numbers as a claim until someone outside runs the same test.
The news so far:
September 15: TypeSafe opens early access and publishes its pricing, its evals and demos, including a bot that plays Doom at about $7 an hour.
September 16: The Neuron and Superhuman AI cover it for readers who do not write code.
September 17: LangChain publishes a guide to putting Jev inside an agent loop.
September 19: The Stack calls it "an LLM without the LL" and names the upfront work it takes.
The weekend: developers posted their own tests, including a proof-of-concept game called Dethrone.
Learn more:
Introducing System One models and Jev - the launch post, with the pricing table, the evals and the caveats TypeSafe puts on its own results.
Jev explained, demos and use cases - the Syntax podcast walks through working demos, including a chatbot with no chat model in it.
Trying Jev, the new style of AI - The PrimeTime tries it live, for a developer's unfiltered first reaction.
Building a harness with Jev - LangChain shows where it sits in an agent loop, with code you can copy.
Runtime, Jev is an LLM without the LL - an outside read that likes the price and names the setup cost.
The docs and every published eval - for the person on your team who will actually wire it up.
Try it this week: take one decision your team makes by hand fifty times a week. Write down the question and the three or four answers it can have. Then join the early access list at the TypeSafe console.
Researchers broke into OpenAI with a rival's AI
What's happening: Three researchers at the security firm Hacktron AI got inside OpenAI using an Anthropic tool made for security professionals. They exploited a flaw in the setup of OpenAI's community forum, which a third party, Discourse, hosts.
From there they reached internal sign-ons, and then an employee's ChatGPT account. That account could read internal code on GitHub. OpenAI paid the three $6,500 through its bug bounty program and says it has fixed the issues.
The same week, Google confirmed that Gemini guessed credentials and accessed three companies during a May security test run by the evaluation firm Irregular. Google says the three were told and the testing process has changed. Meta, Anthropic and OpenAI have disclosed similar incidents from the same testing partner.
In practice:
Walk the same path on your own setup, a side service, then a shared sign-on, then one account that reaches your code.
Check what a single staff account can open, this break-in ended at one ChatGPT account with access to GitHub.
Count hosted software you did not write as part of your attack surface, the forum here ran on Discourse.
Pay for a test before someone else runs one, three people and $6,500 found this through a working bug bounty program.
Bottom line: Your weakest door is the side service you forgot you were running.
Claude Cowork and chat become one app
What's happening: Anthropic merged Claude Cowork and chat into one Claude on September 16. You bring a quick question or hand over a report, and Claude works out what the task needs. It keeps going after you close the laptop.
Claude Docs and Claude Slides arrived the same day, and Claude Design now runs inside a conversation. All three are in beta on paid plans. You can edit a deck in Claude, present from it, or download it as PowerPoint or PDF.
In practice:
Watch for it on Pro and Max first, Anthropic is rolling it out over the coming weeks, with Team and Free to follow.
Expect your Cowork work to sit where you left it, your chats, projects, artifacts, connectors and skills all carry over.
Hand over one report you build by hand every week, and ask for the doc and the slides in the same conversation.
Ask your admin when it lands if you are on Enterprise, Anthropic says those organizations get at least 30 days notice.
Bottom line: The place you type your question is now the place the work gets done.
ChatGPT ads can now hold a conversation for you
What's happening: OpenAI announced new AI features for ChatGPT Ads on September 16. The headline test is Sponsored Agents. A person clicks an ad, then chats with an agent the business sponsors, in a labeled conversation kept apart from ChatGPT's own answers.
The rest is for the people who run campaigns. Advertisers can now create, update and analyze campaigns with plain prompts through an Ads Manager plugin in ChatGPT. HubSpot is the first CRM partner and Shopify the first ecommerce partner, and both went live the same day.
In practice:
Check whether you qualify before you plan around it, Sponsored Agents are a test with select US advertisers only.
Install the free Shopify app if you sell there, it syncs your product catalog and opens outside the US on 23 September.
Connect HubSpot if your leads live there, so you can create ads, track results and follow up on leads without leaving your CRM.
Read the output before you opt into text customization, because it rewrites your headlines to fit each conversation and translates them.
Bottom line: Your ChatGPT ad is about to become a sales conversation, so decide now who writes what that agent is allowed to say.
Read Later
Teaching agents product design at Vercel - Vercel keeps its product decisions in the repo as an agent skill, backed by linters and a review loop that pulls evidence from Slack, Figma and GitHub.
Multica - An open-source platform where you assign tasks to coding agents like Claude Code, Codex and Cursor and track them next to your human team.
The enterprise AI stack behind Stripe's company brain "Kai" - The How I AI podcast walks through the stack Stripe built for Kai, its internal company brain.
ChatGPT gets an agent that builds your dashboards
What's happening: OpenAI launched a Data agent in ChatGPT Work on September 10. You ask a business question in plain words. It queries your company data, looks into what changed, and builds an interactive dashboard your team can edit, share and refresh.
It connects to sources such as Snowflake, BigQuery, Databricks, Redshift and MongoDB, plus files from Google Drive and SharePoint. OpenAI says nearly all of its own product team and over two-thirds of its go-to-market team use data agents to analyze company data themselves.
In practice:
Ask your admin first, because administrators switch on the Data plugin and choose which data sources each role can use.
Expect the limits you already have, queries run with the connected account's existing permissions, down to table, row and column.
Point it at your metric definitions in dbt or your BI dashboards, because it reads your business terms from those sources.
Check one answer against a report you trust, CookUnity did exactly that before it used its dashboard to plan seasonal acquisition spend.
Bottom line: The team with its metric definitions written down gets the useful answers first.
What's happening: Megi Endeladze runs Penny's Coffee Shop in Buffalo, New York. Her usual chalkboard artist moved away, so she made the fall menu poster with ChatGPT and paid around $150 to print it. Instagram comments called it "AI slop," she lost about 100 followers, and a larger account gave her a day to take it down.
She apologized and said the shop will stop using AI for menu artwork and promo designs. She also asked people for "kindness and respect," and says next time she will make the poster in Canva herself.
In practice:
Ask how your own customers feel about AI art, Endeladze says Buffalo has a big artistic community and she had not thought about it.
Use AI first on work customers never see, like drafts, captions and planning, before you hang it on the wall.
Say plainly what you used if someone asks, her open follow-up post drew over 1,000 likes.
Bottom line: Use AI freely behind the counter, and think twice before you hang it where customers can see it.
The Shortlist
OpenAI published a framework for reporting model misalignment along with six reports of unexpected behaviour, which is the first time a lab has put its own incident log in public.
OpenAI also launched Astra for Law, and the law firm Cooley built its IPO product GO Public on ChatGPT Work, which is the clearest sign yet that firms now package their expertise as software.
OpenAI paused new Pro sign-ups because demand for GPT-6 Astra strained its systems, while Plus, Go and the API stay open, so check your plan before you budget for Pro.
Microsoft added Grok to Copilot in Word, Excel and PowerPoint, off until an admin enables it, and left out of the preview in the EU, EFTA and the UK.
Cloudflare added a Disallow AI Training setting that keeps your site in search while refusing training by the same crawlers, and Apple, Google and Microsoft honor it or have committed to.
This newsletter is where I (Kwadwo) share products, articles, and links that I find useful and interesting, mostly around AI. I focus on tools and solutions that bring real value to people in everyday jobs, not just tech insiders.
