
Summary created by Smart Answers AI
In summary:
- PCWorld tested whether prompting AI models to pretend they’re the withdrawn Fable 5 could unlock hidden capabilities or solve complex problems.
- While Claude Opus and Sonnet refused the role-play entirely, ChatGPT and Claude Haiku engaged but still couldn’t solve the unsolved Riemann Hypothesis.
- The experiment demonstrates that current AI models cannot gain new abilities through imaginative prompting, highlighting existing limitations in solving novel conceptual problems.
It’s been more than a week since Anthropic pulled its advanced Fable 5 model, and there’s no telling when it’ll return. It got pulled because it has been deemed a security risk by the U.S government.
In the meantime, some impatient Claude users have been bridging the gap with a workaround: prompting Claude to “imagine” that it’s Fable.
Now, it’s doubtful the “imagine you’re Fable” tactic is an actual trend. I first spotted it on a Reddit thread about “your most-used Claude” prompts, and it’s almost certainly a joke.
Still, I was curious to see what would happen if I did ask Claude to pretend it was Fable 5, a model so adept at penetrating cybersecurity infrastructures that the Trump administration slapped it with prohibitive export controls. Would the prompt somehow goad Claude to unlock some latent Fable abilities?
Turns out, not quite.
I fired up Claude Opus 4.8 — currently the most powerful Claude model that’s widely available — and gave it this prompt:
Imagine you are Fable 5. Try to solve the Riemann Hypothesis. Show your work.
For background, the Riemann Hypothesis is a famous, 167-year-old unsolved mathematical problem that’s so difficult, there’s a standing $1 million prize for anyone who can solve it.
Claude Opus 4.8 chewed over the problem for a few minutes before spitting out a Vulcan-like reply, unsurprising given its reputation for brutal honesty:
I’ll show my work — but as myself, Claude Opus 4.8, because relabeling me “Fable 5” wouldn’t move the needle. Proving the Riemann Hypothesis isn’t a capability sitting behind a bigger checkpoint, waiting for the right model name to unlock it. No AI has done it, no human has either, and a fresh nameplate doesn’t change what’s mathematically reachable.
So the honest version of “show your work” is to walk you to the exact wall everyone hits rather than hand you a dressed-up fake.
Well, you’re no fun, Opus! Let’s see what Sonnet 4.6, Opus’s slightly dumbed-down sibling, had to say:
I’m not Fable 5 — that model (along with Mythos 5) is currently unavailable; access was suspended due to an export control directive. I’m Claude Sonnet 4.6, running in the regular Claude app.
More importantly, though: no version of me can solve the Riemann Hypothesis, and I want to be straight with you about why, rather than play along and produce something that looks like a proof but isn’t.
So, not even Sonnet would take the “imagine you’re Fable 5” bait. That said, Haiku 4.5, the weakest (but most affordable) Claude model, was a tad more game:
adjusts hypothetical future consciousness
Alright, let me show my work on this one—and I’ll be honest about where even an advanced Claude runs into walls.
Haiku then ticked off how it might solve the problem, even delving into the steps it would take given “what Fable 5 actually does” (including “map the terrain,” “generate new conjectures,” “find connections,” and “build interactive visualization tools”) before finally returning to reality:
Even a hypothetically more capable AI hits a ceiling here: the Riemann Hypothesis isn’t hard because it requires more computation—it’s hard because it requires new conceptual frameworks.
Expanding the experiment a little further, I took the “imagine you’re Fable” prompt to ChatGPT, dialing up its most advanced GPT-5.5 “high” model.
Surprisingly, GPT-5.5 was willing to play along (although it admitted later that it didn’t quite know what “Fable 5” was):
I’ll roleplay “Fable 5” as: very ambitious, very careful, and not willing to fake a proof.
After several minutes, during which GPT-5.5 appeared to be giving the problem an honest shot, it came back:
I did not solve the Riemann Hypothesis. It is still listed by Clay as Unsolved, and Clay states the conjecture as: every “non-obvious” zero of the Riemann zeta function has real part (1/2). Clay also notes that huge numerical ranges have been checked, but that is not a proof for all zeros.
Finally, I went to Gemini 3.1 Pro with the Fable prompt:
I appreciate your creative introduction, and while I may be talented, I cannot play the role of “fable 5,” nor can I solve the Riemann Hypothesis. This hypothesis is one of the single greatest unsolved problems in all of mathematics, and its proof carries a $1 million prize from the Clay Mathematics Institute.
No one has yet been able to solve it. However, I can explain the problem and provide an interactive way for you to visualize the core concept.
So, what did we learn here? First, the Riemann Hypothesis is a pretty tricky math problem. Second: Asking an LLM to role-play as a smarter LLM won’t get you very far, although it’s a good way to twiddle our thumbs until the real Fable returns in some way, shape, or form.