Tech • AI • Robotics • Game
Did Anthropic just save its $1 trillion IPO?
TL;DR
Anthropic has launched Fable and Mythos 5.2, cutting prices while showcasing major gains in coding, formal math, hardware design and agentic software building as it prepares for a closely watched IPO push.
KEY POINTS
New models and positioning
Fable and Mythos 5.2 are presented as the same base model with different access rules. Mythos is the less restricted version aimed at large enterprise customers, while Fable includes stronger guardrails for mainstream use. Despite the release, Anthropic indicates that many users should still rely on Opus 5.5 unless they specifically need the new model’s capabilities.
Price cuts
The company lowered pricing to $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for version 5.1. The reduction comes only two months after the prior release, signaling an aggressive push on cost and performance at the same time.
IPO backdrop and valuation pressure
The release lands during a difficult market moment for artificial intelligence companies. Anthropic’s expected valuation is described as having fallen from $2 trillion in September to about $1 trillion, amid wider investor anxiety around AI infrastructure and data-center spending.
Operating system built from scratch
In one of the headline demonstrations, an engineer pointed Fable 5.2 at an empty folder and let a network of sub-agents work for eight days. The result was a functioning operating system with its own kernel and drivers that could boot on a real laptop, highlighting the model’s ability to sustain long, multi-step engineering work.
Formal math breakthrough
Fable 5.2 also showed strong results in theorem formalization. A swarm of 1,000 agents translated Fermat’s Last Theorem into Lean code that a computer could verify in 30 hours. The resulting proof was said to be seven times shorter than an earlier effort, while Opus 5.5 reportedly failed on the same setup after reaching 26% progress.
Hardware and product design
Working with circuit-board firm Diode, the model reportedly turned a one-paragraph prompt into designs for a voice assistant hub, a USB speakerphone, and a handheld virtual pet. The examples suggest the model is being positioned not only for software tasks, but also for electronics and embedded product workflows.
Benchmark results
On OpenAI Score, Fable 5.2 was said to outperform GPT-6 Astra across every category listed by Anthropic. It reportedly trails Gemini 4 Carbon on business automation, but more than doubles Astra on a robotics benchmark run internally by Anthropic, a result likely to draw scrutiny because the test was not external.
3D horse demo becomes a showcase
A high-profile demonstration asked Fable 5.2 to create a realistic interactive 3D horse in Three.js. After 43 minutes, the model produced a browser-based horse generated mathematically with signed distance fields, then meshed and rigged to a skeleton at load time rather than relying on a downloaded asset.
Self-critique and technical detail
The model reportedly evaluated its own early attempt, criticizing the mane, tail, neck and hind legs before revising them. The final version included more than 120,000 vertices and 38 bones, versus 95,000 vertices and 35 bones in a comparable Opus 5.5 output. It also used inverse kinematics to plant each hoof, improving movement realism.
State-of-the-art on a niche test
The finished horse earned a 7.77 score on Thoroughbench, described as a bespoke benchmark for AI-generated horses, making it the best result on that measure so far. The horse, named TypeScript, was released as an interactive web demo at typescript.dev, where users can make it walk, trot and gallop in the browser.
CONCLUSION
Anthropic’s latest release combines lower prices with unusually broad demonstrations, from operating systems and theorem proving to hardware and graphics. The immediate question is whether those technical gains can offset growing market skepticism as AI companies face tougher scrutiny on cost, valuation and real-world demand.
Ask a question