Skip to content

Vibe Research II: My Three Days with Fable 5

Key Ideas

  • Fable 5 is Anthropic's most powerful model — released briefly for Claude subscribers, then shut down within days by US government order after Amazon researchers jailbroked it. (Mythos 5 is its higher-capability sibling, available only to vetted security researchers.)
  • Working with Fable felt qualitatively different — "a level up by an order of magnitude." It thinks two or three steps ahead, is better at execution, and found closed-form solutions to mathematical proofs that four prior models (including Opus 4.6–4.8 and GPT 5.5 Pro) and the author himself had only solved numerically.
  • Used for: comprehensive review of an ongoing paper (catching issues Opus+Codex missed), planning two new research projects (producing first-draft structure from just a research question), solving theory proofs analytically, and generating a complete Manim-based video series (10-minute explainer + 6-part derivation walkthrough) with AI voice-over (Kokoro) in approximately one hour — one-shotting complex multi-file animation projects with minimal follow-up.
  • Token cost: roughly 150% of a 5-hour quota (Claude Max 5x, ~$100/month) for the video work alone, or ~$200 at API prices. Rate was "mind-blowing" — >50% of a 5-hour quota consumed in under 15 minutes during parallel video rendering.
  • The government shut-off raises concerns about differential access: foreign nationals (including US-based faculty and graduate students) are effectively locked out even if they can pay.
  • The author's underlying paper ("Investing in Artificial General Intelligence") was written end-to-end using agentic AI, scored 10th of 159 at the Human × Finance conference, and has ~1,300 SSRN downloads — fully AI-assisted research, fully disclosed.

Summary

Grégoire's "Vibe Research II" documents three days of access to Claude Fable 5 before the US government ordered its shutdown — the second entry in his personal series on AI-assisted academic finance research. Where his first post ("Vibe Research") described writing a paper in four days with earlier models, this post grapples with a qualitatively different capability jump.

Fable 5 caught subtle issues that multiple prior frontier models missed in manuscript review, found closed-form proofs where numerical solutions were the best prior result, planned full first-draft research projects from a research question alone, and produced a professional Manim animation series almost entirely autonomously. Grégoire describes the experience as analogous to first encountering ChatGPT — "trouble sleeping not from stress, but because my brain was processing all the implications."

The post ends on a geopolitical note: the forced shutdown of Fable and Mythos 5 — at US government request after a jailbreak by Amazon researchers — means that foreign nationals, including those at US universities, are cut off regardless of their willingness to pay. This raises questions about differential capability access that will shape international AI research competition.

Relevance to Economics Research

This post illustrates the frontier of what a single researcher can accomplish with a sufficiently capable AI model in a short time window. For economics and finance researchers, the key findings are: (1) frontier models now differ qualitatively, not just incrementally; (2) theory work — proof-finding, analytical solutions — is now within reach of AI assistance in a way it wasn't with prior generations; (3) multimodal, automated output production (video explainers, animations) is now a realistic research communication tool; (4) the cost and access structure matters — Fable 5 at $200 for a session's worth of video work represents a substantial but not prohibitive research investment. The forced shutdown also raises policy questions relevant to economists studying AI governance and international technological competition.