Claude Fable 5.2 Generates Playable Games in Hours, but the Real Costs Add Up Fast

When Anthropic engineer Felix Rieseberg fed Claude Fable 5.2 a single instruction to construct a first-person shooter, the model ran unsupervised for eight hours, consuming 445 million input tokens and generating 2.7 million output tokens to produce a working browser-based game. The final bill came to approximately $550 through the API, though the actual computational resources required tell a different story.

The resulting game, called Saltline, places players in a rain-soaked port environment where they confront drones and robotic dogs while destroying communications equipment. Rieseberg shared the project when Fable 5.2 launched on Monday, framing it as a test of how much the model could deliver from a single directive. The model generated all visuals procedurally without relying on pre-built assets, and the finished product fits into a roughly 1.4MB HTML file that bundles Three.js code. Rieseberg contributed roughly two hours of manual work refining weapon designs, audio effects, and game mechanics while keeping the original single-shot version accessible.

The real price tag

YouTuber Matthew Berman pushed the experiment further by instructing Fable to reconstruct a Gears of War: E-Day level inside Unreal Engine 5 using only three lines of text. The resulting playable shooter allowed players to use cover, fight alongside AI partners, and execute the franchise's signature active reload mechanic. Berman's effort consumed roughly 8.5 billion tokens across multiple days, with an API-equivalent cost approaching $15,000.

Berman also created a browser-based spreadsheet application featuring ribbon tabs, formulas, and pivot tables—a project that required nearly 50 hours and an estimated $11,700 in computational costs. Since Berman had early access to the model, these figures may not reflect what he actually paid, but they offer meaningful insight into the resources required to run agents over extended periods.

Not Fable alone

Several developers credited Fable with creating games that actually relied on multiple tools working in concert. Creative developer Kevin Ngo described his project What the Water Keeps as "Claude Fable 5.2 made this game demo, playable in the browser. It's a boat ride through 8 dreamlike worlds," but Ngo employed Tripo to generate 3D models and Suno v6 to compose the soundtrack, while Fable handled the trailer music using royalty-free samples.

Matt Shumer, co-founder of HyperWrite, invested substantially more time experimenting with the model. He used Fable to develop a browser-based voxel destruction system with physics-based damage, then tackled a more ambitious project: an incomplete simulation of Earth featuring real locations, weather patterns, and transportation networks. Shumer initially tested multiple agents but discovered that a single agent delivered superior outcomes with minimal intervention.

It nailed the voxel engine I've been trying to get out of almost every model for a year

Matt Shumer

Shumer's battle royale combining Hogwarts and Fortnite aesthetics took roughly two weeks to develop. He found that running multiple agents simultaneously produced disappointing results, leading him to adopt a single-agent approach.

It works best as one agent that owns a whole project

Matt Shumer

However, Shumer acknowledged that Fable still requires active guidance. "You still have to steer it. Framing the problem and keeping it on track with your goals is on you," he explained. Shumer hosts his demonstrations on Spawn, a platform listed among his investments, though he did not disclose this connection in his assessment.

The token math

Fable 5.2 charges $8 per million input tokens and $40 per million output tokens, but these published rates obscure the true economics when agents operate continuously. Rieseberg noted that the majority of Saltline's 445 million input tokens benefited from cached context, which Anthropic discounts. Without caching, those input tokens alone would have cost roughly $3,560, compared with his stated $550 total—a dramatic difference that highlights the importance of token optimization.

Beyond raw token costs, developer time remains a significant factor. Someone must decide what to build, examine the generated code, and resolve problems. Shumer had to reconsider his strategy when multiple agents underperformed, while Ngo relied on separate tools to create game assets.

From demo to product

The demonstrations raise substantial questions about what transpires after the initial working version. Future developers must comprehend the codebase, adjust features without introducing regressions, and investigate performance bottlenecks. Berman's spreadsheet illustrates this challenge: formulas that function in a prototype may produce inaccurate results when applied to larger datasets or more intricate calculations.

The projects also prompt inquiry into their origins. Several of the most impressive builds recreate existing games, with Berman replicating a Gears of War: E-Day segment from Microsoft's current franchise entry, and Shumer drawing inspiration from both the Wizarding World and Fortnite for his battle royale.

Peter Gostev, an AI evaluation specialist at Arena, has examined dozens of prompts and found much to praise, though he remains committed to Opus. "It's a great model, a LOT of things got better but not every single thing," he posted on X, noting that one experiment cost $2,618.

Saltline and Berman's creations demonstrate that developers can now delegate substantial implementation responsibilities and return hours later to discover something functional or interactive. Measuring the cost of that work grows increasingly straightforward, yet assessing the quality of the output still demands human expertise.