The world of artificial intelligence continues to amaze: Anthropic's Claude Fable 5 model confidently won a test for realistic physics, but this achievement comes at a cost. Literally. According to data from the analytics service atomic.chat, the cost of generating scenes for Fable 5 turned out to be about six times higher than that of its more modest "colleague" — Claude Opus 4.8.
The essence of the experiment was simple but revealing. Four leading AI models were given the same task: create three independent HTML5 scenes with realistic physics demonstrations. The three episodes involved: a train falling from a destroyed bridge into water, two cars colliding mid-air after jumping off a ramp over a canyon, and a monster truck crushing a row of parked cars. The key requirement was that the physics of falling, destruction, and collisions had to be flawless, without any extraneous artifacts.
Results: Who Leads and Who is Frugal?
Fable 5 completed all three scenes with top marks. The collisions looked realistic, objects fell and broke correctly, nothing fell through the ground or hovered in the air. However, quality came at a price: the model used 62,158 tokens, costing $3.12.
The closest competitor was GPT-5.5. It used 37,753 tokens at a cost of $1.14. In the monster truck scene, according to the service's assessment, it even performed slightly better than the leader.
Model Opus 4.8 proved to be significantly more economical: 22,280 tokens and only $0.56. This is about six times cheaper than Fable 5, but the quality of the scene rendering, unfortunately, was more modest.
The cheapest was GLM-5.2. It didn't win any scenes but cost only $0.08 with 36,246 tokens. The service's overall conclusion is clear: Fable 5 is the best choice for quality, but you have to pay significantly more for it.
Why is Fable 5 Back in Action?
This test came shortly after Fable 5 returned to full availability. As a reminder, on June 30, U.S. authorities lifted export restrictions on Anthropic's Claude Fable 5 and Mythos 5 models, and the restoration of global access began on July 1. The restrictions were imposed on June 12, just three days after the models' release, citing national security concerns. Anthropic was forced to disable both models for clients on Claude.ai, via API, AWS Bedrock, and other partner services, as filtering users by nationality in real-time proved impossible.
Fable 5 is positioned as Anthropic's most advanced and widely available model. It is built on the Mythos-class architecture, with finely tuned safety mechanisms added, and it excels at complex tasks, programming, and image analysis. Internal classifiers automatically redirect high-risk requests to safer algorithms. Model prices remain competitive for the market: Anthropic charges $10 per 1 million input tokens and $50 per 1 million output tokens. This cost was reflected in the atomic.chat test.
My professional opinion: The AI model race is entering a new phase where content generation quality is becoming a key differentiator. Fable 5 demonstrates an impressive leap in physics realism, which is critical for game development, simulations, and visual effects. However, the sixfold price difference with Opus 4.8 is a serious signal for developers: the choice between "good enough" and "perfect" now has a clear price tag. For the mass market, this could be a limiting factor until costs come down. We'll be watching developments closely.