Claude Fable 5.2 tests: demos, benchmarks and pricing

Building AAA-Style Game Levels from Short Prompts

0:15 to 2:15 in the video, about 2 minutes

A major test for Claude Fable 5.2 was recreating a playable cover-shooter level inspired by Gears of War using Unreal Engine. The prompt used was just three lines long, providing minimal explicit constraints while asking for complete graphics, gameplay mechanics, and characters.

The model generated a working level featuring functioning AI enemies, cover mechanics, shooting, and timed active-reload mechanics. Running under maximum reasoning effort, this single project consumed roughly 8.5 billion tokens across approximately 44 hours of build time.

• Generated realistic lighting, cover mechanics, and enemy combatants in Unreal Engine.
• Demonstrated the model's ability to interpret underspecified prompts without breaking gameplay logic.
• Required a massive token budget to compile full 3D game assets and logic.

Replicating Complex Productivity Software

2:16 to 6:03 in the video, about 4 minutes

To evaluate productivity software generation, Fable 5.2 was tasked with cloning Microsoft Excel as a browser-based application and DaVinci Resolve as a desktop video editor. Both applications were completed with functional interfaces and core feature sets.

The web-based spreadsheet tool included interactive formula bars, live calculations, and functional pivot tables. Meanwhile, the DaVinci Resolve clone provided timeline track management, clip slicing, magnetic snapping, and keyframe controls, all operating locally.

• Excel clone was generated in roughly 50 hours, using over 8 billion cached input tokens.
• Spreadsheet output supported real mathematical formulas and pivot tables directly in the browser.
• The video editor replica operated locally on desktop with timeline editing and clip management.

Physics Simulations and Multiplayer Games

6:04 to 10:32 in the video, about 4 minutes

The model proved capable of coding physical simulations from pure mathematical logic. In one prompt, it generated a standalone HTML and WebGL2 water balloon bullet impact simulation without external libraries like Three.js, incorporating customizable bullet velocity, ray-traced shadows, and real-time fluid burst physics.

Fable 5.2 also generated several complete playable games, including a 10-level multiplayer obstacle course inspired by Fall Guys called Tumble Beans, an F-Zero style racer named Starlane, a physics-based puzzle recreation of Portal, and custom mobile ports.

• Raw WebGL2 ray tracing and fluid burst calculation implemented in a single zero-dependency file.
• Tumble Beans featured 10 distinct obstacle levels with functional real-time multiplayer lobbies.
• Included retro arcade racers, platformers, and physics-based puzzle games across web and mobile platforms.

3D Modeling, Architecture, and Large Environments

10:33 to 11:36 in the video, about 1 minute

Testing extended into procedural 3D modeling and large-scale architectural design. Using Blender and Unreal Engine, Fable 5.2 generated fully textured modern villas, functional multi-agent European town simulations with pedestrian dialogue, and physical 3D-printable structures designed to hold weight under strict mass limits.

The largest environment test involved generating an expansive 3D model of Tokyo in Unreal Engine. Complete with skyscraper geometry, moving train lines, street-level vehicle traffic, and ray tracing, the recreation required 6.72 billion tokens across 70 hours.

• Engineered an under-10-gram 3D printable tower capable of supporting multiple laptops and tools.
• Produced fully textured modern architectural house renders with lighting and reflection passes in Blender.
• Constructed a wide-scale urban recreation of Tokyo featuring active traffic and transit lines.

Motion Graphics and Visual Concept Explainers

11:37 to 18:55 in the video, about 7 minutes

Fable 5.2 was tested on generating educational motion graphics with zero post-production editing required. Through single prompt requests, the model created modular graphic explainers showing how compilers deconstruct machine code into assembly and how large language models process embeddings, vector spaces, and token prediction.

Each output included clean visual transitions, consistent styling, and synchronized text callouts, showing high capability in translating complex technical concepts into finished, styled visual presentations.

• Built end-to-end technical explainers on disassembly, control flow, and token predictions.
• Generated consistent typographic styles ranging from whiteboard aesthetics to technical blueprints.
• Produced ready-to-use animation assets in under an hour per topic.

Benchmark Scores and Industry Comparisons

18:56 to 20:34 in the video, about 2 minutes

Official benchmark results show Fable 5.2 securing top scores across standard frontier metrics. On FrontierCode 2.0, the model achieved 49.1%, significantly ahead of competing models, while reaching 69.9% on FrontierSWE V2.

A standout area was OSWorld 2.1, which measures an AI agent's ability to operate an operating system and control a computer directly. Fable 5.2 recorded an 89.3% score, outpacing previous high-end models such as Opus 5.5 and Astra.

• FrontierCode 2.0: 49.1%, leading major frontier alternatives.
• OSWorld 2.1: 89.3%, demonstrating top-tier computer control capabilities.
• Humanity's Last Exam: 74.2%, improving nearly 10 points over prior generations.

Practical Behavior, Reasoning, and Pricing

20:35 to 23:04 in the video, about 3 minutes

Beyond benchmarks, practical workflow testing revealed distinct behavioral traits. The model demonstrated strong autonomy, consistently selecting the optimal architectural path during extensive code reviews and quickly catching subtle edge-case bugs, UI defects, and security flaws.

Anthropic adjusted standard pricing down by roughly 20% compared to Fable 5.1, landing at $8 per million input tokens and $40 per million output tokens. A specialized Fast Mode offering up to 2.5 times generation speed is available at a premium rate.

• Sub-agent orchestration allows extensive reasoning across large codebases in parallel.
• Autonomous decision-making reduced back-and-forth prompt iterations during code refactoring.
• Standard pricing set at $8/$40 per million tokens, with Fast Mode priced at $16/$80 per million tokens.

Final Thoughts

Claude Fable 5.2 establishes a notable step forward in generative code and procedural media creation. While generating large-scale environments and complete software suites remains computationally demanding and token-heavy, the model demonstrates the ability to manage end-to-end execution with very little manual intervention.

As developer tools incorporate more autonomous multi-agent reasoning, the gap between simple prototyping and production-grade software continues to narrow rapidly.