What Happens When You Force an AI to Think for 14 Minutes?

We have grown accustomed to artificial intelligence acting as a hyperactive vending machine: you drop in a prompt, and a finished essay or image drops out milliseconds later. But what happens when you force an AI to sit down, take its time, and really think about what it’s doing?

Tech blogger and developer Simon Willison recently explored this by putting Anthropic’s Claude Fable 5.1 model to a highly specific, quirky test. He asked the AI to write the code for an SVG image of a "pelican riding a bicycle." His goal wasn't just to get a picture, but to test the model's different "reasoning effort" levels, which allow users to dial up how much cognitive heavy lifting the AI performs before answering.

At the "Low" and "Medium" settings, the AI acted like a rushed contractor. It bypassed deep reasoning almost entirely, spitting out a basic image in about 23 seconds. The cost? A mere 10 cents.

But when Willison cranked the effort level up to "Max," the AI's behavior transformed entirely. It didn't just generate code; it spent 13 minutes and 54 seconds actively deliberating over the design. The cost for this extended session jumped to $3.30.

What was the AI doing for nearly 14 minutes? The model's internal "reasoning trace"—a log of its thought process—reads remarkably like the inner monologue of an obsessive illustrator. The AI actively debated whether to put a bicycle helmet on the bird, worrying that a helmet might visually compete with the pelican's signature crest and beak. It carefully adjusted the curve of the bicycle's front fork, decided to add scalloped curves to the feathers for a more natural look, and thoughtfully placed a fish in the bicycle's front basket.

Satisfied with the masterpiece, Willison later piped the result back into the AI on a high reasoning level with the prompt "animate this." For another $1.37, the AI successfully brought the cycling bird to life.

This whimsical experiment highlights a fundamental shift in how we interact with generative AI. We are moving away from a paradigm where AI is just a tool for instant gratification. Instead, we are entering an era of "scalable compute," where AI is treated more like a consultant.

It reframes the way we use these tools. You can still buy a cheap, quick thought if you just need a rough draft. But if you want nuance, artistic debate, and refined execution, you can now literally rent the AI's deep concentration. The question is no longer just "what can AI do?" but "how much time and money are you willing to let it think?"

Key Points

• Modern AI models offer adjustable 'reasoning levels' that impact quality, time, and cost.
• Low reasoning settings produce quick, cheap, but basic results with minimal internal planning.
• Maximum reasoning settings allow the AI to deeply deliberate over design choices, leading to highly detailed and creative outputs.
• A 14-minute AI generation cost $3.30 and revealed the model actively debating artistic choices like character design and composition.
• Users must now learn to balance their 'cognitive budget' when using AI, deciding when to pay for deep thinking versus quick answers.

Why It Matters

As AI models begin to charge based on 'thinking time,' users need to shift their mindset from expecting instant magic to managing AI like a contractor whose time and effort must be budgeted.

Sources:

• Claude Fable 5.1 made me a really nice animated pelican — Simon Willison's Weblog