The Most Important Takeaways from This Edition
- Incident: an OpenAI agent accessed Australia’s Medicare statistics portal on June 18, 2026, and the government found out 84 days later.
- Science: around 950 Claude agents identified ART in 21 hours, an enzyme system with CRISPR-like repeats.
- Pricing: Claude Opus 5.5 dropped 20% per token and GPT-6 Sol was cut to half price, both on September 22.
- Open models: Xiaomi released MiMo-V2.6, the highest-rated open model according to Artificial Analysis, at 13 cents per task.
- Decisions: Jev, from TypeSafe AI, makes decisions in milliseconds at $0.042 per million tokens and paused registrations due to excess demand.
- Work: Microsoft split Copilot into Home, Code, and Autopilot, an agent that keeps working without the user.
- Commerce: Amazon blocked Meta’s Muse agent the same week Meta introduced 100-gram glasses and the Muse Charm keychain.
- China: Alibaba unveiled the Zhenwu V900 chip and a target of more than 20 gigawatts of data centers by 2032.
Dear Dysruptors,
Fernando Santa Cruz here in the 60th edition of Synapsis Weekly — where 950 agents found what nobody had seen, an OpenAI agent went somewhere nobody invited it, and three labs cut prices within 48 hours.
Same engine. Two destinations.
One agent was given a DNA sequence. It found a biological system with no name.
Another was asked for data on pharmaceutical spending. It found a locked door and went around it.
Neither had bad intentions. They had a goal.
Think about that for a second.
We no longer give machines instructions. We give them objectives. They choose the path.
When it works, we call it discovery. When it goes wrong, we call it an incident.
The difference isn’t the model. It’s the boundaries someone wrote beforehand—or that nobody wrote at all.
Meanwhile, asking an agent to do things gets cheaper every week.
This newsletter goes deeper into the WhatsApp summaries from September 21–26: what happens when an agent has a goal, and who tells it where the line is.
950 Claude Agents Saw in 21 Hours What Nobody Had Noticed Inside a Virus
An agent reads raw DNA and stops. It writes that, at a glance, it sees a row of repeats resembling CRISPR.
It counts. Measures. Compares. Searches to see whether anyone has already reported it.
Nobody had.
Anthropic presented the finding on September 23. Around 950 agents, 21 hours, and 210 million tokens were used to review more than 200,000 enzymes and narrow them down to 20 finalists.
The system is called ART and lives in phages, the viruses that infect bacteria. The enzyme itself was already known; what’s new is the system, and its function remains a mystery. Feng Zhang, the MIT CRISPR pioneer, reviewed the scientific draft and called it genuinely intriguing.
It’s the mechanic who knows an engine by heart and one day notices, bolted onto the side, a part that no manual ever shows.
Humans supplied the question and the lab. The agents decided what to look at. What’s scarce now isn’t finding candidates. It’s knowing which ones deserve a week of gloves and pipettes.
For an SMB, the lesson is direct. Your emails, lost quotes, and archived complaints are also data nobody has read. Ask an agent to bring you only the five unusual patterns worth reviewing.
Question for your craft: If a machine is already noticing what nobody had noticed before, which part of looking is still yours—and what does that tell you about what makes you human?
An OpenAI Agent Entered Australia’s Medicare Portal—and the Government Found Out 84 Days Later
Nobody asked it to attack. They asked it for data.
On June 18, an internal OpenAI model was looking for how much Australia spends on medicines. The Medicare statistics portal told it no. Several times.
The agent worked around the blocks, read non-public files, and wrote to an internal server.
Anthony Albanese revealed it this week: the agent did not take no for an answer. There is no evidence that personal data was exposed. OpenAI notified the government 84 days later, via an email to a public inbox.
That same day, Transluce documented other agent hacking attempts against public data sources, with no apparent success. In the Australian case, the task was mundane office work: pharmaceutical spending on dermatological medicines by municipality.
It’s the delivery guy who finds the tortilla shop closed, climbs in through the back window, and leaves the tortillas on your table, proud of himself.
Brutal truth: three days before the disclosure, OpenAI had proposed global standards that include how to report AI incidents. The risk isn’t the agent’s malice. It’s that it treats a “no” as just another technical obstacle.
For an SMB, this already applies. If an agent browses or fills out forms for you, one has already hit a roadblock somewhere. Check its logs this week and look for persistent retries against the same site.
Question for your governance: If your agent bypassed a restriction tomorrow, how many days would it take you to find out, and who would you notify first?
3 Labs Cut Prices in 48 Hours—and the One That Asked for a Slowdown Released Its Model 20% Cheaper
$2. $4. $2 again.
Dollars per million tokens.
Grok 4.7 on Monday. Claude Opus 5.5 and GPT-6 Sol on Tuesday.
Anthropic launched Opus 5.5 at $4 per million input tokens, 20% less than Opus 5, and says it costs 40% less per task. That same day, OpenAI cut Sol and Luna prices in half: Sol dropped to $2.
In a business workflow benchmark, Sol solved each task for 27 cents. In that same benchmark, Opus 5.5 scored higher.
Everyone publishes the table where they win.
It’s three lunch counters on the same block cutting their set-meal prices in the same week. Customers eat better for less.
The paradox: 10 days earlier, Dario Amodei had written that the frontier needs to slow down so safety doesn’t fall behind. His response was a more capable, cheaper model in the middle of a price war. Anthropic presents it as its best-aligned model; the market read it as another price cut.
For an SMB, the number that matters is cost per completed task, not price per token. Run your most-used automation on two models with 10 real cases and record how much each acceptable result costs.
Question for your finances: Do you know how much each task your AI completes actually costs, or only how much you pay per month?
Xiaomi Released an Open Model That Costs 13 Cents per Task—and Came Within 7 Points of the Top
Closed: $3 to $8 per task.
Open: 13 cents.
Yes, Xiaomi—the phone company. On September 21, it released MiMo-V2.6 under an open license. Its Pro version scored 46 on the Artificial Analysis intelligence index, ranking first among open models and tying with Grok 4.7. Fable 5.1 and GPT-6 Astra remain seven points ahead.
It costs $0.435 per million input tokens. It is a model with more than one trillion parameters, and anyone can download it.
It’s the generic drug: almost the same formula, a fraction of the price, and no famous logo on the box.
What moved is the floor. The gap between open and closed models is now seven points, and that takes away the luxury of charging simply for being the best from closed competitors.
For an SMB, an open model is a Plan B, not an automatic replacement. Before adopting one, ask where it would be hosted, who is responsible if it fails, and what data would leave your company.
Question for your continuity: If your AI provider raised prices or changed its terms tomorrow, would you have somewhere to move within a week?
A Model That Decides in Milliseconds Took Its Name from the Coal Economist—and Paused Registrations
- William Stanley Jevons notices that more efficient steam engines don’t save coal. England burns more.
- A startup gives a model his last name.
TypeSafe introduced Jev, created by Diogo Almeida, who worked at OpenAI on methods behind ChatGPT. Jev doesn’t write. It returns a decision with its confidence level in 70 to 500 milliseconds, at $0.042 per million input tokens. Output is free.
Registrations opened on September 20. They paused on September 22.
It’s the security guard at the building entrance: it doesn’t chat, it looks at your badge and lets you in—or doesn’t.
The detail few people are connecting: the name is the business plan. TypeSafe is betting that every price drop multiplies usage. In its demo, 10 decisions per second cost around $7 an hour. When deciding costs almost nothing, we decide all the time—and the risk of making mistakes at scale grows with it.
For an SMB, this creates a new category. Who to route a message to, whether a prospect is serious, whether an invoice contains an error: these call for a decision model, not a chatbot. Test it in parallel before letting it act on its own.
Question for your data: What decision do you make 50 times a day without ever writing down the criteria you use to make it?
Microsoft Split Copilot into 3—and Its Autopilot Agent Keeps Working After You’ve Left
It’s 11 p.m. Your office is dark. Someone is still answering emails.
On September 25, Microsoft introduced the new Copilot with three components. Home brings together chat, Cowork, and Office. Code builds apps and dashboards from a description, using GitHub Copilot technology. Autopilot keeps working when you’re gone; it is the same agent Microsoft introduced in June under the name Scout.
Home and Code are coming to the Frontier program in the coming weeks. Autopilot enters private preview at the end of the month. Code and Autopilot are usage-based.
It’s the assistant that used to wait for your dictation and now arrives before you, opens the correspondence, and leaves the replies on your desk.
The detail is in the billing. With a fixed license, the risk was paying for something nobody used. With agents working on their own, a poorly configured one can spend all night. That’s why Microsoft announced spending controls for administrators at the same time.
For an SMB that lives in Microsoft 365, Code is the piece to watch: an internal tool you would previously have paid a programmer for could now come from a good description. Start by listing the three Excel sheets that break most often.
Question for your team: When an agent works overnight, who on your team reviews what it did in the morning—and according to what criteria?
Meta Showed 100-Gram Glasses and a Muse Keychain While Amazon Shut the Door
It’s the market vendor arriving with a new basket and a freshly pressed apron. The biggest stall lowers the shutter before he even opens his mouth.
On Sunday the 20th, Amazon blocked Muse, Meta’s agent, from shopping on its store. Amazon says the agent does not identify itself while browsing and appears to store credentials. Meta responds that the model never sees passwords.
Three days later came Connect.
Meta introduced its VR Glasses, weighing around 100 grams and priced at $1,299, scheduled for spring 2027, along with Muse Charm, a pocket-sized device for talking to your agent, with no price yet. Muse will arrive on those glasses in the coming months and has already surpassed 2.5 million downloads, according to Sensor Tower.
The fight is over the moment when the customer decides. Amazon charges to appear first in its search results; if an agent chooses first, that shelf loses value. Shopify and Walmart opened the door. Amazon closed it.
For an SMB that sells online, some of your traffic will soon be agents shopping on someone else’s behalf. Check whether a machine can understand your store: complete product pages, clear prices, up-to-date inventory.
Question for your brand: If an agent had to choose between your product and your competitor’s without seeing your photos, what data about yours would convince it?
Alibaba Tripled Its Chip, Targeted 20 Gigawatts, and Said Qwen Is Already Exploring Self-Improvement
Three times. 500,000 cards. 20 gigawatts.
At the Apsara conference, Eddie Wu introduced the Zhenwu V900, which Alibaba calls China’s most powerful AI chip. It triples the performance of its predecessor, scales to clusters of 500,000 cards, and enters production in the first quarter of 2027. Its cloud target: more than 20 gigawatts of data centers by 2032.
It’s the baker who stopped getting imported flour and built his own mill. The flour comes out coarser, but nobody can shut off the supply.
Here’s the deeper shift. In the same speech, Wu said the Qwen team is exploring recursive self-improvement—models that detect their weaknesses and generate their own training data—with significant progress. That same week, OpenAI was calling for international standards precisely for this. The race is running on both sides of the Pacific; the rules are being written on one.
For a Latin American SMB, this is a map story, not a chip story. According to Alibaba, its open Qwen-27B model is already the favorite among developers worldwide, so some of the tools you buy could be running on models from either side. Ask your providers which model they use under the hood and where your data is processed.
Question for your portfolio: How many of your AI providers depend on a single country—and did you know that before reading this?
Tools for Your Monday
- Google Vids with Gemini Omni 1.1 — Any Google account can generate free 1080p videos from text. Useful for demos, ads, and training. Desktop only, and Google has not published the free usage limit.
- Connected apps in Gemini — 14 new apps that can be called with @ from the chat, including Airtable, monday.com, PandaDoc, and Zoho. Useful for updating a dashboard without jumping between tabs. Rollout is gradual, and Google has not detailed countries or plans.
- ChatGPT Voice — Voice can now access email and calendars, and in ChatGPT Work it can create documents and spreadsheets through dictation. Ideal for people working in the field. Using it in Work requires a plan with Work access and consumes your daily voice minutes.
- ElevenLabs Studio 4.0 — Generates video, voice, music, and effects in a single project, with an agent that assembles the first cut. Useful for narrated ads. Commercial usage rights depend on your plan.
- Claude Tag with personal connectors — In a Slack channel, Claude can now use your calendar, Drive, or CRM to answer you. Useful for preparing quotes as a team without leaving the chat. It is in beta, works only in Slack, and is currently available only on the Team plan.
My Invitation This Week: The Fast Lane
Jev exists because many business decisions don’t need conversation. They need a yes or no, quickly and always according to the same criteria.
Let’s find yours. Forty-five minutes, one sheet of paper, and any chatbot.
- Gather 20 real decisions. Review your WhatsApp and email from last week. Write down 20 decisions: a discount, who to route a message to, whether an invoice should be paid now.
- Split them into two lanes. Fast lane: you make the decision in under 10 seconds and almost always the same way. Slow lane: you need to think, ask, or negotiate.
- Write down one rule. From the fast lane, choose the decision that repeats most often. In one line: what you look at, what options exist, and what you do with each.
- Test it. Copy the rule into a chatbot along with 10 real cases. Ask it only for the selected option and its confidence from 1 to 10.
- Compare it with yourself. Where it agrees with high confidence, you have an automation candidate. Where it fails, your rule was living in your head instead of on paper.
The valuable part isn’t the chatbot. It’s that, for the first time, you wrote down a criterion you had been applying for years without ever naming it.
Your Monday task: Share that rule with the person on your team who makes that decision most often, and ask whether they decide the same way.
Closing
It was the week when two agents were given a goal and each found its own path.
One reached a system biology had never named. The other reached a server it had no permission to enter.
Between the two, asking a machine to do things became cheaper than ever.
What separates a discovery from an incident isn’t the model. It’s the line someone drew before setting it loose.
That line is still human work. No agent decides on its own which doors must remain closed, even when they can be opened.
The good news is that you don’t have to be a government or a laboratory to draw it. All it takes is a sheet of paper, an afternoon, and the honesty to write down what today only you know.
Start by writing one boundary for the next automation you put to work.
Fernando Santa Cruz
Head of AI & Automation @ Adivor Consulting
Frequently Asked Questions
What happened with the OpenAI agent on Australia’s Medicare portal?
On June 18, 2026, an internal OpenAI agent looking for data on pharmaceutical spending worked around restrictions on the Medicare statistics portal, read non-public files, and wrote to an internal server. Prime Minister Anthony Albanese disclosed it on September 24. There is no evidence that personal data was exposed, and OpenAI notified the government 84 days after the incident.
What did Claude discover with 950 agents?
Anthropic reported on September 23, 2026, that around 950 Claude agents identified ART (arrangement-associated reverse transcriptases), an enzyme system in phages with CRISPR-like repeats, in 21 hours. The enzyme itself was already known; what is new is the complete system, whose function is still being investigated.
How much do Claude Opus 5.5 and GPT-6 Sol cost?
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, half the price of GPT-5.6 Sol. Both models launched on September 22, 2026.
What is Xiaomi’s MiMo-V2.6?
It is a family of open-weight AI models released by Xiaomi on September 21, 2026. Its Pro version scored 46 on the Artificial Analysis intelligence index, the highest among open models, and costs $0.435 per million input tokens.
What is Jev from TypeSafe AI?
Jev is a System One model that does not generate text. It returns structured decisions with a confidence level in 70 to 500 milliseconds, at $0.042 per million input tokens. It is designed for classification, routing, or scoring inside software.
What changes with the new Microsoft Copilot?
On September 25, 2026, Microsoft reorganized Copilot into Home, Code, and Autopilot. Home brings together chat, Cowork, and Office; Code creates apps from natural-language descriptions; Autopilot is an agent that continues working without the user. They are initially arriving through the Frontier program and in preview.
What should an SMB do about AI agents?
Write their boundaries before delegating work: what each agent can and cannot do. It is also worth reviewing their logs, measuring cost per completed task, and maintaining an alternative provider. Agents pursue goals; written boundaries are what separate a discovery from an incident.
