Anthropic released the Fable 5.2 and Mythos 5.2 language models today. The models have been added to the llm-anthropic 0.31 package, allowing users to run prompts through them via the local llm toolchain.
How the models were tested
The tester provided installation steps and an example command:
- uv tool install llm
- llm install llm-anthropic
- llm keys set anthropic (paste your API key here)
- llm -m claude-fable-5.2 "Generate an SVG of a pelican riding a bicycle"
They ran the same SVG generation prompt on Fable 5.2 across the five available reasoning levels: low, medium, high, xhigh, and max.
Token exhaustion on the "max" setting
According to the test, most levels produced images; the best result came from the xhigh setting. The "max" setting, however, failed to return an image: the model did not complete the request because it ran out of tokens after spending 128,000 internal "thinking" steps.
This behavior mirrors a bug previously observed with Opus 5.5. It is noteworthy because it is new for the Fable series—both Fable 5 and Fable 5.1 were previously able to produce complete images on the "max" setting.
Why this matters
The issue indicates that the highest resource mode (max) may not scale reliably for some workloads and can terminate due to token exhaustion, resulting in failed or incomplete responses. This affects larger, multi-step generative tasks such as complex image generation or prompts that require multiple rounds of internal reasoning.
Next steps
The incident is documented in the context of the llm-anthropic 0.31 integration and is comparable to the earlier Opus 5.5 failure. Users relying on the "max" setting for critical tasks should exercise caution: test lower settings (for example, xhigh) or monitor token usage until Anthropic provides an official fix or explanation.
The test commands and conditions reflect the llm-anthropic 0.31 setup.
