How a single prompt consumed 90% of a daily AI token limit

When a journalist asked Claude AI to build a Breakout-style game, the request consumed nearly all of the free tier's daily token allowance in a single five-minute session.

The incident highlights how quickly generative AI tools can burn through usage limits, even for seemingly simple tasks.

Lance Ulanoff, a tech writer, recently tested Claude's stability by prompting the AI to create a playable game called "Gearbreaker" — a Breakout clone where bricks are replaced by marching robots.

The prompt was detailed but not exhaustive, describing levels, robot materials, and controls. Claude returned a fully functional game in about three minutes.

But the cost was steep: "Claude just informed me that this Breakout task used up about 90% of my credits for the day," Ulanoff wrote.

That figure refers to the free version of Claude's Sonnet 5 model, which operates at a "Medium" processing level.

While Anthropic does not publicly disclose exact token limits for free tiers, users typically receive a daily cap that resets.

Consuming 90% of that cap on a single prompt means the user effectively exhausted most of their generative capacity for the day in under five minutes.

For comparison, the same journalist had spent hours guiding Claude through an Asteroids clone a year earlier, suggesting that model improvements have made prompts more resource-intensive even as they become more capable.

The game itself — named "Gearbreaker" — includes two levels, procedural scaling, and local high-score saving. Ulanoff noted that the output was "stunningly complete" and required no iteration.

Yet the token consumption raises a practical question for users: as AI models grow more powerful, will they also become more expensive to operate on free tiers?

One way to read this is that the convenience of a single, detailed prompt comes at the cost of daily utility — a trade-off that casual users may not anticipate.

This development does not suggest that AI is wasteful or inefficient. Rather, it points to a growing gap between user expectations and the resource economics of generative AI.

For users relying on free tiers, a single creative project could effectively shut down access for the rest of the day.

For Anthropic and similar providers, the incident underscores the challenge of balancing model capability with usage limits that keep services accessible.

Ulanoff's experience is not an outlier. Many users have reported hitting token limits after complex prompts, especially when the AI generates large outputs like code, images, or lengthy text.

The "Gearbreaker" case is notable because the output was a complete, shareable artifact — a sign that the model allocated substantial compute to produce a polished result.

Future users may need to weigh the ambition of their prompts against the remaining balance on their daily quota.

Original reporting sourced from Lance Ulanoff, TechRadar.