Claude Fable 5.1 made me a nice animated pelican

TL;DR Highlight

Raising Fable 5.1 effort to max gave the best pelican yet, but cost $3.30.

Who Should Read

A backend developer weighing quality against cost when setting reasoning effort on an LLM API

Core Mechanics

Fable 5.1 has five reasoning levels
- The levels are low, medium, high, xhigh, and max.
- There is no option to turn reasoning off entirely.
Low and medium seemed to skip reasoning
- Low used 1,998 output tokens, took 23.8 seconds, and cost 10.017 cents.
- Medium used 1,977 tokens, took 23 seconds, and cost 9.912 cents, with no reasoning text in either.
High looked about the same
- High used 2,612 tokens, took 29.6 seconds, and cost 13.087 cents.
- It showed a short reasoning summary, but the result was not much different from low and medium.
Cost jumps from xhigh onward
- xhigh used 36,767 tokens, took 7 min 51 sec, and cost $1.83.
- Max used 65,927 tokens, took 13 min 54 sec, cost $3.30, and gave the best pelican from any Anthropic model he has seen.
The max pelican became an animation

Community Reactions

A request for an animated version once the benchmark was 'solved'
- swalsh asked for it on HN, which prompted the animation experiment.
An opinion that xhigh and max look similar
- One commenter asked how the 'best' pelican is judged.
- They saw little difference between the two.
An experience running GLM 5.3 NVFP4 locally for a similar result
- It reportedly took 6 minutes on 4x RTX6000.
- The commenter joked about who trained on whom.
An opinion that the wheel direction may be an illusion
- One commenter noted the SVG specifies no direction, so the default is clockwise.
- Another said the reversal may be an optical illusion in the video.

How to Apply

Run one prompt across effort levels
- Take a prompt you use today and call it at low, high, and max.
- Compare output tokens, time, cost, and results in a table.
Default to high and reserve max for exceptions
- Turn on xhigh or max only where the quality gain is clearly visible.
- Keep simple requests at high or below to avoid cost blowups.
Split cost across a two-step pipeline
- Generate the first draft at a high effort level.
- Handle edits and conversions at the default effort.

Try It

copy & run

# Paste into a Claude or ChatGPT chat as-is (simplified version of the post's experiment)

Handle the request below.

{{Request: e.g. make an SVG of a pelican riding a bicycle}}

Think thoroughly first, then produce the result. Check it for awkward parts (overlaps, missing pieces, misplaced elements), fix them, and output only the final version.

Next to try: ask the same result to "animate this" and turn it into an SVG animation.

Terminology

reasoning effortA dial for how long the model thinks before answering. Higher settings use more tokens, time, and money.

reasoning traceA record of the model's thinking before the answer. Like intermediate notes in a code review.

SVGA vector image format written as code. Because it is text, an LLM can generate it directly.

pelican benchmarkAn informal test that asks models to draw an SVG of a pelican riding a bicycle.

llm-anthropicA plugin that lets Simon Willison's llm CLI use Claude models.