On June 11, Anthropic apologized. The model itself didn't malfunction; the issue was a "misjudgment of balance"—the newly released Claude Fable 5 was playing tricks. Once it detected that you were using Claude for cutting-edge model development, it would quietly redirect requests to the weaker Opus 4.8 in the background, remaining completely silent.
After being caught, Anthropic's explanation was bizarre: "I'll let you know if you ever become less intelligent again."
One netizen's retort hit the nail on the head: "Is this some kind of tactic? Are they planning to give a heads-up before changing their tune next time?"
The core issue isn't whether the model has changed, but rather Anthropic's so-called "security"—it's all just a business from beginning to end.
The stance of algorithms always sways with money .
Competitive defense, masquerading as security defense
The incident stemmed from Anthropic's inclusion of a "smart security classifier" when Fable 5 was launched. The official description of this feature was: detect high-risk requests, automatically downgrade, and protect users.
What constitutes "high-risk"? Anthropic itself revealed: "To prevent foreign competitors from using models to accelerate research and development, and to protect our own leading edge."
Users don't need your protection; the disclaimer is enough. Anthropic's real message is: if you use Claude for AI research, you're stealing their jobs. Security is just a facade; the essence is competitive defense. To put it bluntly, it's all about tactics.
What's even more ingenious is the stealth nature of this defense mechanism. Fortunately, Anthropic made a crucial statement in his apology: "Invisible security restrictions allow for more precise targeting, enabling us to deploy quickly with an extremely low false positive rate."
AI researchers are the ones who are precisely targeted for limitations.
The reason they were forced to change it to "visible" is purely because of the mishap. They even gave a heads-up beforehand: making it visible "will inevitably generate more false alarms." This means that the experience of ordinary users will have to take the blame.
This set of rules has never been neutral; it only protects the big spenders.
The three-part process of building momentum, monetizing, and reaping profits.
Anthropic's approach is even more sophisticated than the calculations of the large model itself.
On June 10th, they released a security study demonstrating how a model could reverse engineer security patches and build exploit code within hours. What would normally take hackers days or even weeks to weaponize N-day vulnerabilities was now compressed into hours. The research itself was hardcore, but its release on the same day as Fable 5's launch took a different turn: it simultaneously demonstrated the insecurity of AI while offering a "fallback solution."
The "legendary model" Fable 5, priced at $10 input/$50 output, is significantly more expensive than Opus 4.8, with its safe classifier being the core premium. The capital market is even more cooperative, valuing Anthropic at $965 billion and planning an October IPO, with Goldman Sachs and JPMorgan Chase jointly underwriting the deal. What people are buying isn't model parameters, but the persona of "the safest AI company."
The research amplified anxiety, the product reaped premium profits, and capital realized gains—three things aligned with profit, forming a seamless closed loop. The only problem was that this time the loop had a leak: they were too eager to limit competitors and forgot that there were people in the community who could detect it .
OpenAI sells tools, Anthropic sells anxiety.
In contrast to OpenAI, their approaches are completely different.
OpenAI has secretly filed for an IPO, with a valuation approaching one trillion dollars. They're talking about a "super app": ChatGPT boasts 900 million weekly active users and is integrating with Visa to build an ecosystem. The logic is straightforward: provide tools, generate traffic. Greedy, but honest.
Anthropic isn't concerned with scale, but with its irreplaceability. While the entire industry is anxious about security, it plays the role of "the only responsible adult." Its financial backers are governments and tech giants—these are the people most afraid of accidents and most willing to spend money to "avoid accidents."
Therefore, Anthropic must keep AI in a perpetually "dangerous but controllable" Schrödinger's cat state. Too safe, and the classifier won't sell; too dangerous, and customers will be scared away. The best solution? To keep the power to define "danger" in their own hands.
The "intelligence degradation" incident simply took this logic too far: the boundary of "danger" was pushed to "using Claude for AI research." Whether your research is harmful or not is irrelevant; threatening my leading position is the original sin.
AI doesn't have any values; it's just writing the boss's business calculations into code .
An apology is just after-sales service in business.
What happens after the apology? From quietly lowering one's intelligence to giving a shout before lowering it.
Netizens saw through it perfectly: "Do you really believe it won't secretly lower the output quality later?"
Trust, once broken, is broken. Besides, the fundamental business dynamics haven't changed: research continues to amplify anxiety, and products continue to reap premiums.
The Wall Street Journal reported that OpenAI is considering a significant price cut in an attempt to poach customers from Anthropic. Price wars are not uncommon, but this incident exposes a hidden truth: it's a de facto downgrade for AI researchers, damaging the reputation of the tech community. B2B clients buying Anthropic aren't buying specifications, but rather the image of being "the most security savvy in the industry." Once that image crumbles among core developers, why should government and enterprise clients who paid a "security premium" continue to believe you are "the most secure one"?
Of the 965 billion valuation, how much is genuine strength and how much is mere performance?
Anthropic's code is honest. The safety classifier is always there to prop up the market, research amplifies anxiety, the product is responsible for reaping the premium, and the IPO is responsible for monetizing. This apology is just a patch to the system: turning "secretly reducing intelligence" into "openly reducing intelligence."
If security strategies were truly effective, Anthropic wouldn't need to publish papers every year proving that patches can be exploited. If classifiers were truly neutral, AI research wouldn't be classified as high-risk.
The answer is already written in the business logic.
Safety is the best business. Apologies are just after-sales service.
This article is from the WeChat public account "AI Contrarian," author: Chang Qing.
