Je teste Fable 5.2, le meilleur modèle jamais publié

Je teste Fable 5.2, le meilleur modèle jamais publié

Title history
  • Je teste Fable 5.2, le meilleur modèle jamais publié — observed on 20261111-2115 UTC (YouTube)
  • Fable 5.2 expliqué et testé en 15 minutes (sans hype) — observed on 20261110-0839 UTC (YouTube)
🎙 Erwan | IA 👥 18K 📅 November 10, 2026 ⏱ 14 min 👁 20K 📄 expert opinion 🧭 2026-11-10
Available in: English (current) Français

Keywords

Fable 5.2Anthropicbenchmarkblind testtoken cost

Summary

In this video, Erwan introduces Fable 5.2, Anthropic’s latest frontier model, claiming it surpasses all publicly available models. He explains that Fable 5.2 and Mythos 5.2 share the same architecture, differing only in safety guardrails. Key benchmark results include a 64.9% score on Robot Dojo (vs 39.4% for Opus 5.5) and a 41.3% on the Remote Labor Index (vs 21.3%). The model also formalized a proof of Fermat’s Last Theorem using 1000 agents in 30 hours. Pricing is $8 per million input tokens and $40 per million output tokens, lower than Fable 5.1 but higher than Opus 5. Anthropic recommends starting with Opus 5.5 for most tasks, using Fable only when needed. The presenter performs a blind test comparing Fable 5.2, Opus 5.5, and GPT-6 Astra: given a photo of Manhattan, each model must geolocate the exact time and reconstruct a 3D scene with time travel. Fable 5.2 produced the most thorough analysis and the best 3D visualization, while being more expensive and slower. The test is a one-shot, so results are indicative. The video concludes that Fable 5.2 offers exceptional performance but requires careful consideration of task cost and necessity.

191 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides significant added value by combining official announcement data, independent benchmark figures, and a novel hands-on blind test. The argumentation is structured: it first establishes technical superiority through objective metrics, then discusses pricing nuances (per-token vs per-task), and finally demonstrates a concrete, complex application. The presenter openly acknowledges limitations such as the single-shot nature of the test and the variance between models. The reasoning is sound and avoids hype, focusing on practical implications.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high: the video cites Anthropic’s official announcement, links to Claude support documentation, and references Artificial Analysis benchmarking. The personal test is well-designed (blind, same prompt, no metadata) and transparently reports costs and token usage. Corrections noted in the description show attention to accuracy. The title accurately reflects the content, and the ’no hype’ promise is kept. The presenter also provides context by comparing with previous models and discussing when to use each model.

167 words

Title / Content Match

Title perfectly matches content: explanation of Fable 5.2, pricing, and a hands-on blind test, all delivered without hype.

Quality & Reliability

8/10

The video relies on official Anthropic data, third-party benchmarks (Artificial Analysis), and a controlled personal blind test. The presenter acknowledges the single-shot limitation and transparently reports costs and tokens. Corrections are clearly indicated.

Key Moments

Cited Sources

Contribution & Novelties

The original contribution is the blind test comparing Fable 5.2 against Opus 5.5 and GPT-6 Astra on a complex geolocation and 3D reconstruction task, which is rarely performed. The presenter also contextualizes pricing per task rather than per token, offering practical guidance for users.

Pour aller plus loin :

  • Fermat’s Last Theorem — Formal proof is a major AI achievement; here used as a benchmark for formalization.
  • Remote Labor Index — Scale AI’s metric for measuring AI performance on remote work tasks; not directly available on Wikipedia, but Scale AI is a known company.
  • Robot Dojo — Anthropic’s environment for evaluating robotic control from camera images; likely accessible via Anthropic research page.

112 words

Radar Profile

The radar profile shows high scores in quality, technical depth, and reliability, with slightly lower quantity of information due to the short duration. The dominance of qualitative and technical dimensions underscores the video's focus on rigorous testing and expert insight rather than breadth of topics.

Reliability 8/10