Fable 5.2 in llm-anthropic: max mode burned 128,000 tokens and returned no answer

When a new model comes out, a developer's first question is simple: how do I try it quickly? Simon Willison, the author of the llm utility, answered it on release day. Anthropic released Fable 5.2 and Mythos 5.2, and they are already available in version 0.31 of the llm-anthropic plugin. Along the way he found an unpleasant quirk that anyone planning to run the model at maximum settings should know about.

How to try it in a minute

llm is a command-line utility that provides a single interface to different language models. The llm-anthropic plugin connects Anthropic models to it. To run Fable 5.2, you need four steps: install llm via uv, install the plugin, save your Anthropic API key, and run a command with the model name claude-fable-5.2.

For testing, Willison uses his signature test: asking the model to draw an SVG image of a pelican riding a bicycle. It is convenient because the task requires both spatial reasoning and clean markup, and the result is immediately visible to the eye. Of course, this is not a scientific benchmark, but it is a quick indicator of how well a model handles an unconventional task.

Five reasoning levels and one failure

Fable 5.2 has five reasoning effort levels: low, medium, high, xhigh and max. The higher the level, the more tokens the model spends thinking before it answers. Willison ran the prompt at each of them.

xhigh turned out to be the best. In the image, the legs are on the correct side of the bicycle, the frame is drawn well, and the beak with its pouch looks convincing. There are flaws too: strange little fingers on the handlebars, and the pouch is slightly offset relative to the upper part of the beak. Overall, the result is good.

max, however, returned no image at all. The model spent 128,000 tokens on reasoning and hit the limit without ever reaching an answer. In effect, it overthought the task until it broke.

An important detail: this is not the first such case. Willison observed the same thing with Opus 5.5. But for the Fable line it is news. Fable 5 and Fable 5.1 handled the max level fine and returned images without problems. That suggests a regression in the behavior of version 5.2 specifically, although it is too early to draw final conclusions from a single test.

What to do about it in practice

First takeaway: the maximum reasoning level does not mean the best result. It is more expensive, slower, and, as we see, may produce no answer at all. If you are building a product on Fable 5.2, start with high or xhigh and go higher only where it measurably improves quality.

Second takeaway: always build in handling for an empty response. If the model has exhausted its token limit on reasoning, your code should be able to recognize this and retry the request with a lower effort level, rather than crashing or showing the user a blank. Cost deserves separate attention: 128,000 tokens wasted is real money, especially with high-volume calls.

Third takeaway: when moving to a new model version, run your own regression tests on all the modes you use. What worked in 5.1 may behave differently in 5.2.

Limitations

This is one person's observation on one task. We do not know how often the failure reproduces on other prompts, and Anthropic has not commented on it yet. The problem may be resolved by adjusting limits or by an update. But as a signal for caution, this is enough: test max on your own tasks before enabling it in production.