The market for large language models (LLMs) continues to surprise not only with quality but also with pricing policies. In a recent comparative test on realistic physics, the Anthropic Claude Fable 5 model confidently outperformed its competitors, demonstrating impressive simulation of complex scenarios. However, this superiority comes at a cost—literally.
Test Details: Physics Without Compromise
Four leading models—Claude Fable 5, GPT-5.5, Opus 4.8, and GLM-5.2—were tasked with creating three independent HTML5 scenes demonstrating realistic behavior of physical objects. The scenarios included: a train falling from a destroyed bridge into water, two cars colliding mid-air after jumping from a ramp over a canyon, and a monster truck crushing a row of parked cars. The key requirement was the absence of artifacts and correct simulation of falls, destruction, and collisions.
Results: Fable 5 Leads, But the Price Tag Shocks
Fable 5 handled all three scenes with top scores. Objects fell and broke correctly, without falling through textures or hanging in the air. However, quality came at a price: the model consumed 62,158 tokens, and the task cost $3.12.
The closest competitor, GPT-5.5, showed comparable results, using 37,753 tokens and costing $1.14. Notably, in the monster truck scene, GPT-5.5 even slightly surpassed Fable 5 in simulation quality.
The Opus 4.8 model, on the other hand, proved significantly more economical: 22,280 tokens and just $0.56. This is roughly six times cheaper than Fable 5, although the visual quality of the scenes was more modest.
The most budget-friendly was GLM-5.2—$0.08 with 36,246 tokens—but it did not win any scenes, showing the worst result.
Context: Fable 5's Return After Restrictions Lifted
It is worth noting that the test was conducted shortly after Fable 5 became fully available again. On June 30, U.S. authorities lifted export restrictions on the Anthropic Claude Fable 5 and Mythos 5 models, which had been imposed on June 12, just three days after their release, due to national security concerns. Anthropic was forced to disable both models for clients on Claude.ai, via API, AWS Bedrock, and other partner services. Global access was restored on July 1.
Analytical Conclusion
Claude Fable 5's victory in the physics test confirms its status as Anthropic's most advanced model, built on the Mythos architecture with enhanced protection mechanisms. However, the current cost of $10 per 1 million input tokens and $50 per 1 million output tokens makes its use economically justified only for tasks where quality is critical. For mass applications, where the balance of price and performance matters, Opus 4.8 or GPT-5.5 appear to be more rational choices. The AI model market is clearly moving toward segmentation: premium solutions for complex tasks and cost-effective options for routine operations.