A three-line prompt, a playable shooter #
Matthew Berman, demonstrating Anthropic’s Claude Claude Fable 5.2, showcased a fully playable Gears of War clone built in Unreal Engine. The demo featured dynamic cover mechanics, enemy combat AI, and a signature timed reload system, all generated from a concise prompt. Even a player revive sequence was included, turning a broad request into a functional game level.
Crucially, Berman’s prompt was intentionally underspecified: "Build the most accurate clone of Gears of War E-Day using Unreal Engine. You only need to build one level, but it should look and feel as complete (graphics, gameplay, characters) as the real Gears of War E-Day that just was released." The model’s ability to independently fill in missing design decisions, from enemy behaviors to environmental details, proved more striking than merely generating code.
This ambitious run, however, came with significant computational costs and a notable caveat. Berman pointed out an awkward cover animation, a minor imperfection in an otherwise impressive output. The project consumed 8.5 billion tokens over several days in "ultra code mode," incurring an estimated API cost of approximately $15,000. This was not instant game development, but rather a multi-day, resource-intensive autonomous build.
It didn’t stop at games #
Beyond games, Claude Claude Fable 5.2 demonstrated a startling breadth of application. Berman showcased a browser-based Excel-style clone with fully functional formulas and pivot tables, completed in approximately 50 hours. Another demo featured a desktop video editor, paralleling commercial software like DaVinci Resolve.
The AI also generated a Fall Guys-style multiplayer game called Tumble Beans, featuring multiple levels and complex physics. This ambitious project took about 71 hours and carried an estimated API cost of $13,000. In contrast, a smaller, configurable 3D water-balloon simulation was built for roughly $91.
These figures underscore a crucial distinction: while impressive, the projects requiring extensive compute time and tokens also incurred substantial costs. Berman’s demonstrations and cost estimates highlight the potential for AI to generate complex software.
However, these are the creator’s reported runtimes and estimates. While the feature lists are extensive, they do not constitute independent verification of full parity with established commercial software suites. The capability to build is clear; the cost-effectiveness and functional equivalence across all features still invite further scrutiny.
The real breakthrough is the long run #
Beyond the impressive individual demos, the true paradigm shift lies in the extended build-test-revise cycle. We're moving past asking an assistant for a single function to directing an autonomous agent through a multi-day process that generates substantial, complex artifacts. This is less about code snippets and more about orchestrating a sustained, goal-driven development effort.
A human role remains crucial, as seen in the video. The user sets initial goals and constraints, evaluates gameplay mechanics and user interfaces, spots flaws like the "awkward" barrier reach, and ultimately decides whether to publish or iterate. This is a shift from direct coding to orchestration and critical evaluation, guiding the AI through refinement loops.
While the provocative claim "software is dead" makes for good headlines, these demos suggest lower barriers to rapid prototyping, not necessarily the end of software engineering. They offer little insight into the long-term challenges of:
- Security vulnerabilities
- Maintainability of complex codebases
- Accessibility for diverse users
- Dependable production use
These demonstrations from Anthropic’s Claude Claude Fable 5.2 hint at a future where the initial heavy lifting of software creation is profoundly accelerated. For more on the underlying models, visit the Anthropic Official Website. The human role evolves from artisan to architect, overseeing the automated construction of increasingly sophisticated digital worlds.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
The price tag changes the story #
A compelling prototype is not yet a product customers can trust. The Claude Claude Fable 5.2 demos, while spectacular, come with significant caveats regarding cost and practicality. Running for dozens of hours and consuming 8.5 billion tokens for the Gears of War clone or 50 hours for the Excel clone translates to an estimated $15,000 in API costs alone. This compute budget is feasible for a showcase but prohibitive for routine development by most builders.
Beyond the initial build, a finished product demands much more. Critical requirements like robust testing, intricate bug fixing, third-party licensing, performance optimization, and ongoing maintenance remain largely unaddressed by these autonomous agent systems. These are complex, iterative processes that currently still depend heavily on human oversight and specialized skills.
Therefore, the real measure of these systems isn't the most dazzling demo or the bold claim that developers are obsolete. Instead, evaluate them by the useful work completed per dollar and per human review hour. This practical lens separates aspirational hype from shipping reality, grounding the incredible potential in tangible, economical value for everyday builders.
Frequently Asked Questions #
What does the video show Claude Fable 5.2 building?
Matthew Berman demonstrates projects including a playable Gears of War-style level, an Excel-like spreadsheet, a video editor, a multiplayer game, and 3D scenes.
How much did the Gears of War-style demo cost?
Berman estimates about $15,000 at API pricing after roughly 8.5 billion tokens. That is his estimate for the demo, not a general model price.
Does the video prove AI has replaced software developers?
No. It shows that an AI system can produce ambitious prototypes with sustained runs, but does not establish production reliability, maintainability, or lower total costs.
What is the main limitation highlighted by the demos?
The projects can require enormous token budgets and long runtimes, and some still have visible flaws. Human review remains important.