Anthropic Fable & Mythos 5.2: Speed, Cost, and 3D Horse Benchmark
Anthropic has released Fable and Mythos 5.2, new AI models that promise significant improvements in speed and cost efficiency compared to their predecessors, which were released just two months prior. This launch comes as Anthropic prepares for its IPO roadshow, amidst a challenging market environment where AI stock valuations have seen a downturn.
Fable and Mythos 5.2: Key Features and Capabilities
Fable and Mythos are essentially the same underlying model, with Mythos being the unrestricted version intended for large enterprises with robust cybersecurity measures, and Fable incorporating guardrails to prevent misuse.
Pricing and Recommendations
The new models offer reduced pricing: $8 for input and $40 for output, down from $10 and $50 for version 5.1. However, Anthropic suggests that users primarily stick with Opus 5.5 for most tasks, only switching to Fable when its specific capabilities are truly needed.
Impressive Demonstrations
The new models have showcased remarkable abilities in various complex tasks:
- Operating System Creation: An engineer pointed Fable at an empty folder, and after eight days of autonomous work by sub-agents, it produced a functional operating system. This OS boots on a real laptop and includes its own kernel and drivers.
- Mathematical Proofs: A swarm of 1,000 Fable agents successfully converted Fermat's Last Theorem into verifiable Lean code in just 30 hours. This proof was seven times shorter than previous attempts. In contrast, Opus 5.5 failed to complete the task, crashing at 26% progress.
- Hardware Design: Collaborating with circuit board company Diode, Fable designed a voice assistant hub, a USB speakerphone, and a handheld virtual pet from a single-paragraph prompt.
- Benchmarking Performance:
- On the TMBBBS, Fable 5.2 outperformed GPT-6 Astra across all metrics listed by Anthropic in OpenAI's scoring.
- It was surpassed by Gemini 4 Carbon in business automation tasks.
- Fable 5.2 more than doubled Astra's performance on a robotics benchmark conducted by Anthropic.
The "Thoroughbench" Challenge: A Realistic 3D Horse
To truly test Fable 5.2's capabilities, it was tasked with creating a realistic, interactive 3D horse using 3JS, a benchmark known as "Thoroughbench."
The Creation Process
Fable spent 43 minutes on the task. Instead of using pre-existing models, it sculpted the entire horse from mathematical signed distance fields. These fields are then meshed and rigged to a skeleton each time the page loads.
The model even critiqued its own initial attempt, noting issues like a bluish mane, stick-like tail, and thin neck and hind legs, before autonomously correcting them.
The Result: "TypeScript"
The final output is described as the most realistic horse ever seen on Thoroughbench.
- Visuals: Excellent skin texture and lighting.
- Interactivity: The horse's head follows the mouse cursor, and it can be steered with arrow keys or orbited in 3D. It exhibits shyness when approached too closely.
- Movement: Transitions between gaits (standstill to gallop) are smooth, though stopping can be a bit awkward.
- Technical Details: Fable's horse features over 120,000 vertices and 38 bones, compared to Opus 5.5's 95,000 vertices and 35 bones. It uses inverse kinematics for realistic hoof placement, avoiding simple mesh warping.
- Minor Flaws: The model did not hallucinate a low-poly phallus like Gemini, but it did equip the horse with rear leg shock absorbers, which are visible during rearing. The mane was slightly lopsided. Fable also admitted that the horse was not photorealistic and its eyes were "half-litted."
Score and Naming
Fable 5.2 achieved a score of 7.77 on Thoroughbench, setting a new state-of-the-art for horse generation. Given the increasing sophistication of these models, the created horse has been named "TypeScript," in homage to the legendary racehorse Secretariat.
Users can interact with TypeScript at typescript.dev, where they can walk, trot, and gallop the horse directly in their browser.
MX: Video Infrastructure and AI Workflows
The video also highlights MX, a sponsor that provides a highly customizable API for integrating video features into applications.
MX Features:
- Easy Integration: Simplifies adding video capabilities.
- Automated Processing: Upon video upload, MX automatically generates transcripts, storyboards, thumbnails, and clips.
- AI Workflows (MX Robots): Offers AI-powered tools to identify key moments in videos or select optimal thumbnails without requiring users to host models or maintain pipelines.
- Directives: Users can define workflows once, and MX Robots will automatically apply them to all new uploads, keeping up with the latest AI models.
- Trusted by Industry: Companies like Perplexity, Patreon, and others utilize MX.
MX offers a free plan that includes 100,000 robot units and delivery minutes per month, with an additional $50 credit available at mx.com/fireship.
Takeaways
- Anthropic introduced Fable and Mythos 5.2, models that are up to twice as fast and cheaper than the previous 5.1 release, with input priced at $8 and output at $40.
- Mythos is the unrestricted enterprise version, while Fable includes safety guardrails, and Anthropic still recommends using Opus 5.5 for most workloads unless Fable’s unique abilities are required.
- Demonstrations showed Fable autonomously building a full operating system, converting Fermat’s Last Theorem into Lean code, and designing hardware prototypes from a single paragraph prompt.
- In the “Thoroughbench” test, Fable generated a highly detailed, interactive 3D horse named “TypeScript” in 43 minutes, scoring 7.77 and surpassing previous state‑of‑the‑art results.
- The video also highlighted MX, a video‑infrastructure platform that automates transcription, storyboarding, and AI‑driven thumbnail selection through its MX Robots API.
Frequently Asked Questions
What is the Thoroughbench challenge and how did Fable 5.2 score on it?
The Thoroughbench challenge is a benchmark that asks an AI model to generate a realistic, interactive 3D horse using 3JS, requiring on‑the‑fly geometry creation and animation. Fable 5.2 completed the task in 43 minutes, sculpting the horse from signed distance fields and earning a score of 7.77, a new state‑of‑the‑art result.
Why does Anthropic advise using Opus 5.5 for most workloads instead of Fable?
Anthropic advises most users to stay with Opus 5.5 because it delivers a balanced mix of speed, cost, and reliability across typical tasks, while Fable’s higher performance is tied to specialized capabilities that may not be needed and can incur additional guard‑rail overhead. Switching to Fable is recommended only when its unique functions, such as autonomous code generation, are essential.
Who is Fireship on YouTube?
Fireship is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.
Does this page include the full transcript of the video?
Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.
Helpful resources related to this video
If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.
Links may be affiliate links. We only include resources that are genuinely relevant to the topic.