Web Experts has used Claude since it came out, and the relationship has been a cycle of changing plans, usually downgrading out of protest or frustration. The subscription pays for Anthropic's frontier model, Fable 5.1, but actually getting to use it feels like finding the pot of gold at the end of a rainbow.
Basic programming tasks get flagged constantly. Every safety alert silently kicks the session down to an older Opus model, with no notice and no explanation.
Silent Model Switching
The bigger worry is not the downgrade itself, it is what the downgrade does to the conversation. If a chat is bouncing between models in the middle of a thread, what happens to its memory of what has been worked on? Each model has its own way of handling context, and that difference is not cosmetic.
We have watched Claude lose track of decisions made earlier in a thread, ask questions that were already answered, and contradict its own previous output. There is no proof that model switching is the cause, but the timing lines up well enough to be uncomfortable. Anthropic does not tell you when a switch happens or what context carries over.
A tool that quietly changes brains in the middle of a task is not one we can trust with long-term work. That is not a philosophical objection, it is a practical one about billable hours.
Three Days Of Hand-Holding
Long-term work is exactly what we tried. We ran a continuous, three-day agentic task to see what the model could handle without supervision, and the honest answer is: not much without constant hand-holding.
On the fast setting, it would veer in the wrong direction and burn an absurd number of tokens on work that had nothing to do with the goal. It never examined the visual output it was producing. It failed to keep the long-term objective in view and repeatedly missed details in the data that a careful human would have caught within minutes.
The amount of shepherding required is the part nobody in the "agents are here" crowd wants to discuss. Autonomy sounds great until you are the one babysitting it for seventy-two hours.
The Doom Talk Reads As Positioning
Which brings us to Dario Amodei. It is hard to take end-of-the-world warnings seriously when the product behind them cannot complete a multiday task without wandering off, or get through a Python script without a safety intervention.
We do not think the doom talk is sincere. It reads like positioning: make the technology sound dangerous enough to justify gatekeeping it, pique public interest, and stall open-source AI development in the process. That is a business strategy dressed as a moral stance.
A Pause Is A Fantasy
Even granting that the concern is genuine, the calls for a slowdown are not realistic. A real pause would require governments that do not cooperate on much of anything to suddenly cooperate on this one thing. Anthropic knows that, and everyone in the industry knows it too.
We still hope Fable 5 gets unshackled, because the underlying model clearly has potential when it is not tripping over itself. Right now, though, apocalyptic framing wrapped around a product that cannot build something solid without constant hand-holding says more about marketing than it does about risk.