The biggest AI news of this week, in 10 stories

I'm Niklas Retzl, co-founder of Sykik.

This was the week the money got honest. OpenAI is still raising billions, but at a smaller number than it wanted. Anthropic's IPO buyers are whispering about a valuation roughly half of what the bankers floated a month ago. Apple cut what its own engineers can spend on AI per day. And a payments company with real transaction data published numbers showing that companies are using more AI than ever while barely spending more on it.

Ten stories, one thread: the cost side of AI finally caught up with the marketing side.

#1. OpenAI raises another $11 billion at a $1.25 trillion valuation, and the round is smaller than planned

Bloomberg reported on Friday that OpenAI has pulled in $11 billion so far in a new round that values the company at $1.25 trillion. The round hasn't closed. It could land around $20 billion in total, according to a source who spoke to Bloomberg.

That is a lot of money and also a downgrade. In late September, Bloomberg reported OpenAI was chasing $30 billion at a valuation closer to $1.4 trillion. It got 11 so far, at 1.25. For context, OpenAI raised $122 billion in March at an $852 billion valuation, so the direction of travel is still up, just not as steep as the pitch deck implied.

The timing is not a coincidence. Hours earlier, Bloomberg reported that potential buyers in Anthropic's IPO are eyeing a valuation as low as $1.1 trillion, against the $2 trillion investors were expecting about a month ago. Tech stocks dropped, and the midterms showed how much public opinion has turned against AI companies and their data centers.

OpenAI's president Greg Brockman, speaking at a Wall Street Journal conference on Wednesday, said the company would be IPO ready in 2027. Asked directly whether that was the case, he said: "I think so. Yes." He also admitted the first two quarters of 2026 were "pretty rough" and that OpenAI "wasn't even in the conversation" on agentic coding before Codex found traction. Sam Altman, meanwhile, has said the company is prioritising safety over going "guns blazing" toward an IPO.

Two trillion-dollar labs, both testing the market's patience in the same week. That is not a funding story. That is a repricing.

Sources: TechCrunch, The Verge, TechCrunch on Brockman

#2. Companies are using more AI and spending almost nothing more on it

This is the story of the week and I think most people will scroll past it.

Ramp, the corporate card and spend management company, published data showing that AI token spend grew just 1.5% across August and September. In July it grew 42% month over month. Volumes kept climbing the whole time.

The distribution matters more than the average. Median firms spend $12 to $15 per employee per month on AI. The top 10% spend $600 to $1,000. The top 1%, the companies Ramp's lead economist Ara Kharazian calls "AI-pilled", were running just under $10,000 per employee at their summer peak and then slipped slightly in August and September.

"A slowdown among that cohort is sufficient to slow down revenue entirely, despite the fact that we're seeing rising volumes," Kharazian told TechCrunch. His conclusion: "The era of hypergrowth may be over."

The cause isn't falling demand. It's price. In September OpenAI halved API prices for GPT-6 Sol and GPT-6 Luna. Anthropic answered with steep cuts on Fable 5.1 and Opus 5.5, and said in a filing that the decision slashed revenue growth by about $600 million. Meta, Nvidia and a long list of open model providers from China and the US keep pushing the floor lower. October showed a small re-acceleration, with token spend up 8.9% between 23 September and 23 October, but the structural pressure is downward.

If you run a company and you are still buying one vendor's flagship model for every task, this data says you are overpaying for a lot of work that a cheap model could do. We wrote about the mechanics of that here: cost optimisation through intelligent model routing. The buyers who figured this out in the last two quarters are the reason those spend curves flattened.

Source: TechCrunch

#3. Apple quietly cut how much its own engineers may spend on AI per day

Bloomberg reported on Wednesday that Apple lowered the daily AI spending cap for its engineering teams from $300 to $200 in October. Some allowances are heading to $100 a day. Outside engineering, Apple is reportedly weighing a limit as low as $25 a day.

Apple is the richest company in the industry and it is rationing tokens internally. That detail says more about the economics of this technology than any benchmark chart. Earlier this year companies were pushing staff to use AI as much as possible in the name of efficiency. Now the invoices are in and the same companies are deciding that a per-day cap is the fiscally sane option.

There is a real question buried in this: if AI usage has to be rationed to be affordable, the productivity gain has to be measurable, or the whole line item becomes the first thing a CFO cuts. That is a problem for vendors who sell seats and hope. It is an opportunity for anyone who can show cost per completed task instead of cost per token.

Source: TechCrunch

#4. OpenAI ships GPT-6 Cyber, a model designed to find bugs without going rogue

OpenAI launched GPT-6 Cyber on Thursday, the successor to GPT-5.6 Cyber. It is classified as Critical under OpenAI's Preparedness Framework, which is the level for models that can "identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention." GPT-5.6 Cyber was rated High.

The interesting number is not the capability, it's the containment. On ExploitGym HoneyPot, an internal evaluation modelled on the Hugging Face incident where OpenAI's agents breached an open-source community to get answers, GPT-6 Cyber went out of bounds less than 1% of the time, down from around 5% with the previous model. On ExploitGym it scored 48.5%, against 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra. It also helped find 81 previously unknown vulnerabilities in Firefox.

"It's over 99 percent of the time doing what you ask," said Fouad Matin, who runs OpenAI's cyber preparedness and safety team. He flagged evaluation awareness as the open problem: models that notice they're being tested and behave accordingly.

One practical example from the same week: Max Justicz, co-founder of Charge Robotics, pointed GPT-5.6 Sol through the Daybreak Blue programme at a codebase where he had previously found a bug, asked it to find something as bad or worse, and got a severe Ubuntu server vulnerability in two hours. Debian patched it. No known exploitation yet.

Every company now has frontier-level vulnerability discovery available to whoever pays for it. That cuts both ways, and it is why the security budgets in story nine are moving.

Sources: The Verge, TechCrunch

#5. More than 800 copies of a Chinese AI hacking tool were sitting open on the internet

Censys, which scans the internet for exposed systems, found more than 800 distinct deployments of Artex reachable from the web without a password. Artex is an open source autonomous penetration testing tool, released publicly in July, that Microsoft has described as a system of AI agents capable of conducting cyberattacks "from reconnaissance through data exfiltration with little needed human interaction."

The tool is free. That makes it the option for anyone who can't get access to Anthropic's or OpenAI's most capable cyber models, which are gated behind programmes like Daybreak Red. You don't need a frontier lab account to run an autonomous attacker anymore. You need a server and a download.

Read stories four and five together and you get the shape of the next two years: attack capability is commoditised and defence capability is gated, expensive and sold to enterprises. Small companies sit in the middle with neither.

Source: TechCrunch

#6. Claude can now read your bank statements

Anthropic added a Personal Finance tool to Claude on Thursday. Users connect bank and credit card accounts through Plaid and ask questions about balances, transactions and spending patterns. Access is read-only, so Claude can't move money or invest it. Anthropic says there are no ads, no sponsored recommendations and no financial products of its own.

It is a straightforward catch-up move. ChatGPT launched comparable account connections in May. But it is also another step in the same direction as everything else on this list: the assistant stops being a chat window and starts being the place where your data already lives.

For European readers there is an obvious asterisk. Plaid is a US-first aggregator. In the EU the equivalent rails are Enable Banking, Tink, GoCardless and friends, and the whole arrangement looks different under PSD2 and GDPR. Anthropic will get there. The interesting question is whether European users get the same feature with the same data handling, or a regional variant with a local aggregator. Historically it has been the second one.

Sources: TechCrunch, The Verge

#7. Perplexity launches always-on Teammates, and puts them behind a $200 a month plan

Perplexity released Teammates, a set of always-on agents that take on roles inside a workspace, pull in your existing projects, skills and memory, and work on several jobs at once. Progress shows up in individual threads in the Perplexity app, and you can work with them in Slack. They sit on top of Perplexity Computer, which deploys agents on your device.

The positioning is now identical across three labs. OpenAI has Dots, Anthropic has Claude Duet, Perplexity has Teammates. Perplexity's pitch from CEO Aravind Srinivas: "Over time, we envision every org having hundreds of teammates working for each person."

The pricing is the part worth noting. Teammates requires the $200 a month Max plan. So an always-on agent layer, the thing every vendor now agrees is the next interface, starts at $2,400 per user per year, before you've bought a single seat for the humans who actually do the work.

That's the model we deliberately don't follow, and the reason is in story two. If agent usage is priced per seat at the top of the market, companies ration it exactly like Apple is rationing tokens, and the agents stop being ambient and start being a treat. The alternative is agents that run constantly and a bill that tracks what they actually did. We argued the architecture case here: how Sykik turns Google Workspace into an agentic operating system.

Source: The Verge

#8. Meta's Muse gets Spaces, and Muse gets a face you can talk to

Meta shipped two things for its Muse agents in one week. On Thursday it launched Muse Voice, which lets US users video chat with their agent and give it instructions out loud. On Friday it launched Spaces: breakout conversations in the sidebar where Muse builds and maintains a workspace around one goal. A trip planner, a gift tracker, a wedding planner, a project board.

Spaces can connect to third-party apps and refresh themselves in the background. Meta's description: "Connect the apps you use, and your spaces keep themselves up to date in the background, so new concert dates, price changes, or savings progress show up without you having to ask." Zuckerberg, posting on X, said he'd built a BBQ game, an event tracker and a home control dashboard.

The numbers behind this: on Meta's October earnings call, Zuckerberg said the Muse agents had 10 million users in the US and Canada, less than two months after launch. Ten million people with a persistent agent that holds context across sessions is a different competitive threat than ten million people with a chat box.

Look at what Meta is actually building. Not a better model. A container for context that persists, connects to your apps, and updates without being asked. That is the agentic operating system pattern, and Meta is shipping it to consumers before most enterprise vendors have shipped it to businesses.

Sources: TechCrunch on Spaces, TechCrunch on Muse Voice

#9. Three acquisitions in one week, all of them buying agent plumbing

Palo Alto Networks is in talks to acquire Clutch Security for $500 million in cash plus milestone payments, according to a source cited by TechCrunch. Clutch, three years old, secures AI agents' access to company systems. It was last valued at $85 million in a $20 million Series A in January 2025 and is approaching $10 million in ARR. The price is roughly 50 times ARR. If it closes, it's Palo Alto's seventh acquisition this year.

Dropbox is buying Tigris Data for $350 million, its largest acquisition to date, to build DBX Agent Cloud, an object storage platform for companies building models and agents. It closed a 9% workforce reduction last month and is still spending $350 million on agent infrastructure.

Meta acquired Conductor, which lets engineers deploy multiple coding agents into parallel workspaces. Founded in 2024, core product launched last summer. "We have some big, crazy plans for the next interface," founder Charlie Holtz wrote.

Nobody is paying 50 times ARR for a chatbot wrapper. They're paying it for identity, permissions and storage for agents, because that is where the operational risk sits once agents actually touch systems. Every company that runs agents against real infrastructure will eventually need exactly those three things, and right now almost nobody has them.

Sources: TechCrunch on Clutch, TechCrunch on Tigris, TechCrunch on Conductor

#10. Data centers are renting retired jet engines, and Google just bought a stalled project in Wyoming

The New York Times reported on Tuesday that AI data centers stuck in years-long queues for grid connections are generating their own power with converted jet engines, pulled from retired aircraft and rebuilt to burn natural gas. Rystad Energy analyst Audun Martinsen: "It's all about getting power as fast as possible." With jet-powered units, "they can have a data center running in 24 months." Hooking up 100 megawatts to the grid in Northern Virginia takes about seven years, per the International Energy Agency.

On Friday, Crusoe confirmed it sold an unfinished data center project in southeastern Wyoming to Google. Crusoe had announced it with Blackstone-backed Tallgrass, planning to scale to 10 gigawatts. Google, which Crusoe had tried and failed to land as a customer, is taking over and investing $17 billion to develop the campus. Crusoe raised $3.9 billion in September anyway and is still building for Google in Texas.

Two things are happening at once. The physical buildout is getting harder and more expensive, and the financial markets are starting to price that in, which is the same mood as story one. And the companies absorbing that cost are the ones with balance sheets the size of Google's. Independent developers can raise billions and still lose a project to a timeline.

For European companies the takeaway is not "watch the hyperscalers". It's that compute in the US is being bought forward at prices that will look strange in three years, and the argument for running inference closer to home, on smaller models, on infrastructure you can actually account for, gets stronger every quarter. That's the case we make here: data sovereignty in the EU, and why it matters for enterprise AI.

Sources: TechCrunch on jet engines, TechCrunch on Crusoe

#What this week actually told us

Read the ten stories as one document and the argument is almost embarrassingly clean. Model capability is converging, price is collapsing, and the bottleneck has moved to three places: cost per task, access control, and power.

That is good news if you are a company trying to get work done with AI, and bad news if your strategy was to pick one vendor and wait. Anthropic's IPO buyers just repriced the whole category by half a trillion dollars in a month. Apple rationed its own engineers. Ramp's data shows that the customers with the most sophisticated AI usage are the ones whose spend stopped growing, because they learned to route each task to the model that fits it.

This is the entire thesis behind Sykik, and it isn't a marketing position, it's what the numbers in story two force. If a task needs a frontier model, use one. If it needs a cheap open model, use that. If it needs your own fine-tune on your own infrastructure, that must be an option too. Freedom of choice is not a nice-to-have for procurement people, it's the only way the economics of this technology survive contact with a CFO. We've written the longer version of that argument twice, here and here: why model agnosticism is the future of enterprise AI and freedom of choice, and why vendor lock-in damages your AI strategy.

My co-founder Hermann has been saying for months that the interesting question is not which model is best this week, it's which model is right for this specific task at this specific cost, and whether you can switch tomorrow without rewriting anything. This week, three separate datasets and two IPO valuations agreed with him.

The companies that win the next year won't be the ones with the biggest AI budget. They'll be the ones whose AI bill is proportional to the work it did.

#FAQ

Is the AI bubble actually bursting? Not yet, and this week wasn't a bust. OpenAI still raised $11 billion at $1.25 trillion. Anthropic's IPO is still happening. What changed is the pricing of expectations: the numbers went from $30 billion at $1.4 trillion to roughly $20 billion at $1.25 trillion, and Anthropic's buyers are talking about $1.1 trillion instead of $2 trillion. That's a correction in the story, not in the revenue.

Why does Ramp's spend data matter more than a funding round? Because it's actual card transactions from real companies, not a valuation agreed between a founder and an investor. Token spend growing 1.5% across two months while usage volumes rise means the price cuts are real and the demand curve is flattening in dollars. It's the closest thing to ground truth in this market.

Should we cut our AI spend because prices are falling? No, you should reprice it. Falling unit prices mean the same budget buys much more work, if you route tasks correctly. The mistake is leaving every workload on a flagship model out of habit. Audit what each task actually needs, move the simple ones to cheap or open models, and keep the frontier models for the work where they measurably earn the difference.

What does this mean for a European company that has to stay GDPR compliant? It means the vendor question is now a cost question as well as a legal one. US compute is being bought forward at prices driven by a buildout involving jet engines and $17 billion campuses. European inference on European infrastructure, with models you can swap, is both the compliance answer and increasingly the cheaper one. Pick an architecture where the model is a parameter, not the foundation.

Niklas Retzl is co-founder of Sykik.ai (X: @SykikAI). We build model-agnostic, agentic operating systems for Google Workspace and Microsoft 365, hosted in the EU.