Anthropic’s Fable 5.2 represents a significant development in AI modeling, demonstrating its capacity to tackle complex and resource-intensive tasks. For example, it successfully created a fully playable Gears of War-style game with advanced mechanics, though the project required up to 12 billion tokens and incurred an estimated $15,000 in API costs. These results suggest that Fable 5.2 is particularly suited for high-budget applications, according to Matthew Berman.
Discover how Fable 5.2 performs in game development, including its ability to produce AAA-level projects like Tumble Beans. Gain insight into its capacity to replicate productivity software, such as a browser-based Excel alternative, and its engineering achievements, including the creation of a 3D printable tower designed to meet strict specifications. This guide examines the model’s strengths, challenges, and practical applications for demanding scenarios.
Anthropic’s Fable 5.2: A New Benchmark in AI Models
TL;DR Key Takeaways :
- Fable 5.2 demonstrated exceptional capabilities in game development, creating AAA-level outputs like a fully playable “Gears of War clone” and a multiplayer game inspired by Fall Guys, with projects costing up to $15,000 and requiring billions of tokens.
- The model excelled in productivity tools, producing a “literally perfect” browser-based Excel clone and a Da Vinci Resolve clone with flawless functionality, though costs for these projects reached approximately $11,700.
- Creative and technical applications included a 3D simulation of a bullet piercing a water-filled balloon, rendered entirely from scratch, and a detailed Blender house model, though minor visual errors were noted in the latter.
- Benchmark testing highlighted Fable 5.2’s performance, achieving an 89.3 on OSWorld 2.1 and outperforming competitors on several benchmarks, while also significantly reducing task completion times with Ultra Code mode.
- Fable 5.2’s high costs, with input tokens priced at $8 per million and output tokens at $40 per million, make it suitable for high-budget, resource-intensive projects, though its limitations and expense may deter smaller-scale users.
Fable 5.2, Anthropic’s latest AI model, has been noted for its exceptional power and versatility. Testing highlighted its ability to perform across diverse applications, including game development and architectural rendering. However, its high performance comes with a premium cost, making it suitable for demanding, resource-intensive use cases.
Game Development: AAA-Level Outputs
Fable 5.2 demonstrated its capabilities in game development with a fully playable Gears of War clone. Built using Ultra Code mode, the game featured advanced lighting, textures, and mechanics like barrier-based cover, timed reloads, and full battles. Focused on a single level, the project consumed 8-12 billion tokens at an estimated API cost of $15,000. Despite a minor animation flaw, the game was described as “fun to play” and executed the requested level flawlessly.
Another notable project was Tumble Beans, a multiplayer game inspired by Fall Guys. It included 10 uniquely designed levels and realistic physics, requiring over 9 billion tokens, 71 hours of build time, and an estimated cost of $13,000. The result was a polished, enjoyable experience with monetization potential.
Additional projects included a Call of Duty-style multiplayer game with block-style graphics, a Portal clone with functional but subpar visuals, and Dragon Rider, which featured high-quality graphics. Starlane, similar to F-Zero, stood out for its visuals, while Tidekeeper, a sequel to a previous game, introduced a crab as the main control element. All narration in these games was AI-generated.
The model also created a large-scale 3D visualization of Tokyo, complete with traffic, pedestrians, trains, and ray-traced reflections. Every building was constructed from scratch by a new cloud model. This project used 6.72 billion tokens, took about 70 hours, and cost an estimated $11,000.
Explore further guides and articles from our vast library that you may find relevant to your interests in Fable.
- Anthropic Reportedly Delays Claude Fable 5.1
- Fable 5.1 Targets a Tentative Late August 2026 Launch
- ChatGPT 6 Astra Defeats Fable 5.1 in 10 of 15 Test Scenarios
- OpenAI Anticipates Public ChatGPT 6 Release Between Sept 3 to 10
- Claude Fable 5.1 Launches with Strict Safety but Less Flexibility
- Claude Fable 5.1 Launches with a 25% Drop in Token Consumption
- ChatGPT 6 Astra vs Fable 5.1: Astra Wins 70% of Visual Design Votes
- OpenAI ‘Bell’ Model Leaks May Preview the ChatGPT 6 Successor
- New Anthropic Claude Fable 5.1 Tops Terminal Bench Science Tests
- Claude Fable 5.1 Adds 75% Repeat Discount but Higher Cache Fees
Productivity Tools: Excel and Da Vinci Resolve Clones
Fable 5.2 excelled in replicating complex productivity software. A browser-based Excel clone was described as “literally perfect,” offering features like formulas, pivot tables, and an example sales tracker. This project required 50 hours, 8 billion input tokens, and 15 million output tokens, costing approximately $11,700.
Similarly, a Da Vinci Resolve clone included essential features such as waveform and video displays, clip splitting, and magnetic snapping. Another video editor, Kestrel, developed for Windows PCs, offered keyboard shortcuts, zoom functionality, and keyframe insertion. Unlike the browser-based tools, Kestrel could potentially be adapted for Mac or web use. Both tools were noted for their flawless functionality.
Creative and Technical Applications
The model’s versatility extended to creative and technical projects. A 3D simulation of a bullet piercing a water-filled balloon was coded entirely from scratch, without external libraries or dependencies. Users could adjust parameters like bullet speed, caliber, lighting, gravity, and balloon size in real time, with physics recalculated instantly. The project, which included a 9 mm bullet as an example, took 2.5 hours and cost $91 for 41 million tokens.
In Blender, the model rendered a detailed house with a pool, solar panels, and realistic lighting. The project used 13 million tokens and cost $30. While visually impressive, minor errors, such as a tree placed in the driveway, were noted.
Motion graphics for technical explanations were also produced, including visualizations of software decompiling and LLM training processes. The software decompiling project, completed in 45 minutes using 9.5 million tokens at a cost of $25, delivered high-quality visuals for complex topics.
Engineering and Design
Fable 5.2 demonstrated engineering capabilities by designing a 3D printable tower that met specific constraints. Using a Bamboo Lab P2S printer with a 4 mm nozzle and PLA filament, the tower weighed under 10 g and withstood significant weight before collapsing. The design process took just over an hour, used 10 million tokens, and cost $37.
Benchmarks and Performance
Benchmark testing validated the model’s performance. It achieved an 89.3 on OSWorld 2.1, surpassing Astra at 79.5 and Opus 5.5 at 87. Other scores included 74.6 on Terminal Bench Science and 74.2 on Humanity’s Last Exam, nearly 10 points higher than Fable 5.1. On the Frontier SWE benchmark, it scored 69.9, outperforming all other models, including Opus 5.5. However, it scored 51 on the Automation Bench benchmark, falling short of Carbon’s 53.
Fable 5.2 also demonstrated remarkable speed and autonomy. Tasks that previously took 60-120 minutes with Ultra Code now required a maximum of 10 minutes. It made correct decisions nine times out of 10, significantly improving workflows. During a full code review using Max Thinking effort and Ultra Code, it identified numerous bugs, security issues, and UI inconsistencies.
Cost and Accessibility
Fable 5.2’s pricing reflects its capabilities. The model costs $8 per million input tokens and $40 per million output tokens, a 20% reduction from Fable 5.1. Fast mode, offering 2.5 times the speed, costs $16 per million input tokens and $80 per million output tokens, though it was not available for testing. Fable resets depend on subscription status, account access, and subscription duration.
Limitations and Considerations
While Fable 5.2 outperformed Opus 5.5 in speed and capability, Opus 5.5 remains more efficient and cost-effective. Ultra Code mode, though powerful, is extremely expensive and unsuitable for most users. Recreations of real-world events, such as the Roadster launch, were less impressive, highlighting occasional limitations in specific contexts.
On balance, Fable 5.2 represents a significant leap in AI capabilities, particularly for high-budget, high-complexity projects. However, its cost and occasional limitations may restrict its accessibility for smaller-scale applications.
Media Credit: Matthew Berman
Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.
