In partnership with
Welcome to OHMO AI. Anthropic and OpenAI both cut what their newest models cost, and Anthropic says its model misbehaves less too. Microsoft shut down a scam service whose chatbot read hacked inboxes and pointed criminals at where a fraud was most likely to succeed.
Meta patched a hidden setting that could send your voice somewhere else. And Alibaba's own AI chip lifted its shares.
In today's newsletter:
Anthropic and OpenAI both cut the price of their newest models
Microsoft shuts down a scam service whose chatbot read at least 12,000 hacked inboxes
Meta patches a hidden Muse setting that could send your dictation to an attacker
Alibaba unveils its own AI chip and plans to expand its data centres to more than 20 gigawatts by 2032
LATEST IN AI
ANTHROPIC & OPENAI
Anthropic and OpenAI both cut the price of their newest models
Image source: OHMO AI
OHMO AI: Anthropic released Claude Opus 5.5 on Tuesday with output tokens priced at $20 per million, down from $25 on the model before it, and says the new model leads the field on coding and knowledge work. OpenAI's new models, trained with methods similar to those behind GPT-6 Astra, arrived at lower prices than their GPT-5.6 counterparts. Anthropic rates Opus 5.5 as comparable to Mythos on biology and cybersecurity, so it ships under the same safeguards as its Fable model.
The details:
Cache reads on Opus 5.5 cost 60% less than on Opus 5, a deeper cut than the 20% on input and output: $4 per million input tokens and $0.20 per million cache reads.
Anthropic says the model "attempted to circumvent boundaries around 85 percent less often" than past models, a figure the company measured against its own earlier releases.
The company also says Opus 5.5 produces writing that is easier to understand.
OpenAI says its new models improve on professional work, on coding and on computer use.
Why it matters: The two cuts read differently once you look at what each is measured against. Anthropic's $20 is a posted price a finance team can write into a contract, while OpenAI's figure is up to 50 percent off promotional pricing, a discount on a discount. The tell will be what OpenAI charges once that promotion ends: if the price holds, both labs moved, and if it climbs back, only Anthropic did.
PRESENTED BY BELAY
The New Rules of Online Visibility
Your customers are searching in places your strategy doesn’t reach.
So before your business is buried and left behind, you need to understand the new rules of SEO.
BELAY's SEO in the Age of AI report explains how search is changing, what AI means for your visibility, and the practical steps small businesses like yours can take to stay visible.
BELAY’s U.S.-based Marketing Assistants turn strategy into execution, helping your business stay visible, credible, and competitive in every search.
MICROSOFT
Microsoft shuts down a scam service whose chatbot read at least 12,000 hacked inboxes
Image source: OHMO AI
OHMO AI: Microsoft said on Tuesday it led an industry-wide disruption of EvilTokens, a subscription scam platform whose AI chatbot could read a victim's inbox and point criminals at trusted relationships, payment approvals and the moments a fraud was most likely to land. The company says at least 12,000 of its customers' accounts were compromised over a few months, belonging to more than 10,000 organisations around the world.
The details:
Working through a legal process and a network of partners, Microsoft seized 50 websites and at least 150 more domains used to run the service.
London's Metropolitan Police arrested at least two men, aged 32 and 38, on September 18, who allegedly ran the EvilTokens website as its administrators.
Microsoft described the tool as an "AI-style chatbot" sitting at the centre of the service, and said it could also flag "sensitive responsibilities" inside an inbox.
The highest concentration of the affected organisations was in the US.
Why it matters: The expensive part of a fraud used to be a person reading an inbox to work out who trusts whom and who can approve a payment. EvilTokens sold that as a subscription feature, and seizing websites does nothing to the chatbot behind it. Two administrators in custody in London is the part of this that lasts; watch what happens when the customers those two served go looking for a new supplier.
META
Meta patches a hidden Muse setting that could send your dictation to an attacker
Image source: OHMO AI
OHMO AI: Meta has patched its Muse app for macOS after security researcher Patrick Wardle found a zero-day that could let someone take control of the AI agent. The app carried an undocumented setting that any code running on the Mac could change without special privileges, sending dictation to an attacker's own endpoint instead of Meta's servers and exposing the spoken audio and prompts along the way.
The details:
The setting is named endo_voyager_dictation_endpoint, per Wardle's project repo, and appears nowhere in Meta's documentation.
Redirecting the transcription traffic also handed the attacker access to the victim's Muse account.
David Singleton of Meta Superintelligence Labs said on X that an attack needs malicious code already running under the user's account, so the practical risk to Mac users "was therefore quite low".
Why it matters: The malware is the ordinary part. The unusual part is that an assistant sold on giving you control came with a hidden switch deciding where your spoken words went, and no screen where you could see it or lock it. Meta has issued a patch for this one switch; what changes for Muse users depends on Meta listing the rest of its settings, because a control promise only covers the ones you can find.
PRESENTED BY HUBSPOT
Watch the full system run live on Sep 30. Walk away ready to do it yourself.
Most founders have LinkedIn traction with nothing to show for it in the CRM. On Sep 30, Maria Gharib (Mindstream) and Valerie Chapman (Ruth AI) walk through the exact system live.
From AI-assisted content creation to sequenced outreach to booked meeting. You'll leave with a process you can run the same day.
Eligible startups also get the LinkedIn-to-Leads Toolkit: ad credits, Apollo, Captions, and HubSpot's Prospecting Agent.
ALIBABA
Alibaba unveils its own AI chip and plans to expand data centres
Image source: OHMO AI
OHMO AI: Alibaba introduced the Zhenwu V900 at its cloud unit's Apsara Conference in Hangzhou, a chip it says delivers three times the performance of its previous Zhenwu M890 processor. It also set out plans to expand Alibaba Cloud's data centre capacity worldwide to more than 20 gigawatts by 2032. The Hong Kong-listed shares closed 2% higher on Tuesday and the US-listed stock gained more than 3% in premarket trading.
The details:
Mass production and commercial release of the V900 are scheduled for the first quarter of 2027.
The M890 it is measured against was released only in May.
Alibaba has not revised its capital expenditure guidance of RMB 380 billion.
The company framed the chip and the data centre plan as part of a broader roadmap covering chips, cloud infrastructure and AI models.
Why it matters: Nobody can buy a V900 before 2027 and every performance figure is Alibaba's own, so Tuesday's move reads like expectation being repriced rather than revenue. The chip doing the work in Alibaba's data centres today is the M890 from May. If three times the performance turns into cheaper computing on Alibaba Cloud's price list when the V900 ships, the claim holds, and if the prices don't move, the shares rose on a roadmap.
AI GUIDES
Stop the AI mixing up your instructions and your pasted text
Paste an email under a question and the AI often summarises the wrong thing, follows an instruction buried in the email, or answers the pasted text instead of you. Put a plain label on each part and it can see exactly where your instructions end and the material begins. It takes ten extra seconds and works in any chatbot.
Copy this, swap in the bracketed parts, and paste it in.
INSTRUCTIONS:
[What you want done, for example: summarise the text below as five bullet points for someone who has not read it.]
BACKGROUND:
[Anything the AI needs to know, for example: this is a client email and I am the account manager replying.]
TEXT:
"""
[Paste the email, article, notes or document here.]
"""The labels and the triple quotes mark out which words are your orders, which are context and which are raw material to work on, so the AI acts on your instructions rather than on anything inside the pasted text. Use the same labels every time and it becomes a habit.
COMMUNITY
Featured Reader
Every newsletter, we feature how a reader is using AI to work smarter, save time, or make life easier.
Today's reader is Sungmin L. from South Korea:
“I work at an apartment block in Ulsan doing maintenance work, fixing the ignition and replacing pumps. Now the part the customer argues about, the bill. Made a Gem in the Gemini app on my phone. Run it in the stairwell before I ring the bell. Type the model, the fault code, the parts, the price, then type “Write a Kakaotalk message to the customer. 3 short lines. What was broken, what I did, why it costs this. Say which part was the money. No sorry, no technical words. Under 200 characters.” Comes back in Korean, 3 lines, customer reads it while I pack up. Paid on Kakaotalk before I'm down the stairs most days. AI took over the annoying part. ”
Want to be in the next newsletter? Tell us how you use AI here.
AI Meme
Go Deeper
AI Guides: Practical AI guides, prompts and tools.
Instagram: The AI page they pray you’ll never find.
Facebook: For the ones who never left Facebook.
