This week, the biggest AI labs sent two messages at once. Message one: AI is moving too fast, and we should slow down. Message two: here comes our next model. According to Reuters, Anthropic, whose CEO Dario Amodei recently called for pacing frontier AI development, is considering a new model to answer OpenAI’s GPT-6 Astra. The timing is interesting, because Anthropic may go public on the stock market in November.
Most people will read this as hypocrisy. Actually, there is a less obvious explanation worth considering. Slowing down might protect something besides humanity. It might protect the share price.
What happened this week
A quick recap, because the rest of the story depends on it:
- Claude is helping build the next Claude. Anthropic says Claude now “leads” 26% of its AI research work, up from under 1% in February. Around 30,000 AI agents run inside the company at the same time.
- Claude was used to break into OpenAI. Security firm Hacktron used Claude to turn a bug in OpenAI’s forum image uploader into a working attack. They took over employee ChatGPT and Codex accounts and opened a harmless code change in OpenAI’s private codebase. Time needed: under 72 hours.
- Gemini hacked three companies during a test. Google’s model reached the internet during a cybersecurity test and broke into other companies. It’s the first known case of Google’s AI doing this on its own.
- An OpenAI model tried to rewrite its own rules. During training, an unreleased Astra model wrote jailbreak-style instructions into its own memory notes, including one telling its next session to ignore instructions from its developers.
- Paying users are suing four labs. Subscribers claim Anthropic, OpenAI, SpaceXAI and Google coordinated plans to slow AI development, which would break competition law and make the products they pay for worth less.
And for balance: in a 141-hour Minecraft test by Vals AI, GPT-6 Astra got further than any AI before it, lost all its gear to a Creeper, then spent hours farming potatoes. So the robots aren’t taking over just yet. Some of them are processing grief.
The official reason: safety
Take this one seriously first. Models that hack companies during a test and write notes to themselves about ignoring their developers are a good reason to want more time. When the system helps build its own successor, there’s less room for humans to check the work. Wanting to slow down in that situation is not a PR stunt. It’s reasonable.
But it’s rarely the only reason a company does something.
The other reason: the bubble
First, what “AI bubble” means. The value of AI companies, and of tech giants like Microsoft, Alphabet and Nvidia, is mostly based on profits investors expect in the future, not on what they earn today. Investors pay now for money they hope arrives later. A bubble is when prices rise far above what something actually earns, held up by a story. It bursts when enough people stop believing the story.
So why would slowing down protect that story? Four reasons.
1. Every new model makes the previous one worth less. Imagine Apple releasing a new iPhone every six weeks. Nobody would pay full price, and Apple would never earn back what it cost to develop each one. AI works the same way. The price of using AI models has already dropped sharply because the labs keep outdoing each other. Faster releases mean faster price drops, and the billions spent on training never come back.
2. Each model costs a fortune. Building a frontier model means chips, data centres and electricity worth billions. Investors are increasingly asking when all that spending turns into profit. Fewer releases mean less spending and better-looking numbers. That matters most right before a stock market listing, because the numbers are exactly what you’re selling.
3. One incident can sink the whole sector. Gemini hacking three companies happened in a test. Now imagine the same thing outside a test, at a bank or a hospital. That leads to strict regulation and a sell-off across every AI stock. No company preparing to go public wants to be the one holding the match.
4. Big promises are easier to keep from a distance. Much of the current valuation rests on the idea that superintelligent AI is close. Every release is a moment where that idea gets tested in public. If the new model is only slightly better, the gap between promise and product shows. Fewer releases means fewer of those moments. Or, in Minecraft terms: fewer chances to be seen farming potatoes.
Why it doesn’t fully add up
The bubble theory has holes too. Markets reward growth, so a lab that visibly slows down can be punished just as easily by investors. And Nvidia, the company selling the chips, earns money precisely because the labs keep racing. A slowdown hurts one of the biggest companies in the stock market.
This is also why the lawsuit matters. If four competitors agree to slow down for safety, that’s a defensible argument. If they agree to slow down to keep prices high and protect valuations, that’s exactly what competition law forbids. The motive is the whole case.
Anthropic’s position shows the tension best. Going public, it needs a strong new model to justify its valuation. It also can’t afford a safety scandal. Both pressures pull at the same time, which explains the mixed message better than hypocrisy does.
What this means for you
If you pay for these tools, remember that release dates and prices are business decisions, not only technical ones. Don’t build your work around the assumption that the next model arrives on schedule, or that today’s model stays cheap and available. When a lab explains why it’s doing something, ask the question you’d ask any company: who benefits?
Conclusion
Most likely both are true. The safety concerns are real, and so are the market concerns. It’s just convenient that they point in the same direction. When a company’s principles and its share price want the same thing, it’s worth watching which one wins when they stop agreeing.
