# Did Anthropic just save its $1 trillion IPO?

## Executive summary

Anthropic released Fable and Mythos 5.2, claiming significant performance improvements over previous versions (e.g., Opus 5.5). The models demonstrated advanced capabilities, including building an operating system from scratch, solving complex mathematical proofs faster than competitors, and generating highly realistic interactive 3D models. These advancements are presented as crucial for Anthropic's IPO roadshow.

## Key takeaways

- Model Differentiation and Pricing: The models are priced at $8 in and $40 out, down from $10 and $50 (5.1). Mythos is the unrestricted version for large companies, while Fable includes guardrails.
- Benchmark Performance (TMBBBS): Fable outperformed GPT-6 Astra across most listed benchmarks. It lost to Gemini 4 Carbon in business automation but more than doubled Astra's performance on a proprietary robotics benchmark.
- 3D Modeling Benchmark (Thoroughbench): Fable 5.2 generated a realistic, interactive 3D horse using 3JS, achieving a score of 7.77 on Thoroughbench. This involved sculpting the animal using signed distance fields and rigging it to a skeleton.

## Technical details

- Operating System Generation: Fable was shown to generate a fairly legitimate OS that boots on a real laptop, starting from an empty folder. This was achieved by a swarm of 1,000 Fable agents in 30 hours, compared to an unreleased Claude taking 11 days for a similar task.
- Mathematical Proof Generation: A swarm of 1,000 Fable agents completed a math project (related to Form's last theorem) in 30 hours, resulting in a proof 7 times shorter than the attempt by Opus 5.5, which crashed out at 26% with the same setup.
- Hardware Design: Fable used a single paragraph prompt to design a voice assistant hub, a USB speaker phone, and a handheld virtual pet in collaboration with a circuit board company called Diode.

## Practical implications

- The rapid advancement of multimodal AI models (like Fable 5.2) suggests a shift toward using AI for complex, multi-step engineering tasks, such as full OS development or advanced hardware design.
- Developers should monitor the performance metrics of new models against established benchmarks (like Thoroughbench) to assess real-world capability.
- The integration of video APIs (like MX) allows for automated content generation, including transcripts, storyboards, and key moment detection, streamlining content pipelines.

## Topics

AI Model Comparison, Generative AI, Build Engineering, Large Language Models (LLMs), Machine Learning Benchmarking, MX (Video API), TypeScript

Source: https://www.youtube.com/watch?v=1FurWDiK9A4
