What happened
Simon Willison’s November 9, 2026 announcement introduces version 0.31 of llm-anthropic, a plugin for the llm command-line tool. Willison says the release adds Claude Fable 5.2 after Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. For people who already use the CLI, the immediate change is model selection: prompts can now be sent using the claude-fable-5.2 model name.
Willison’s setup example begins by installing llm, then installing llm-anthropic, and setting an Anthropic API key. It invokes the model with llm -m claude-fable-5.2 followed by a prompt. His example asks for an SVG of a pelican riding a bicycle. The announcement therefore describes a command-line route to the model rather than a new interface, new API protocol, or a separate hosted application.
To test the model, Willison ran that same pelican prompt at five reasoning levels: low, medium, high, xhigh, and max. He says the xhigh run produced his best result. His accompanying assessment describes an image with the legs on the correct side of the bicycle, a good frame, and a pouch, while noting odd fingers and a slightly misaligned pouch. The claim is qualitative, based on this particular visual task, rather than a scored benchmark across a larger prompt set.
The max setting produced a different operational result. Willison says it failed to return an image at all after running out of tokens, having spent 128,000 tokens thinking about the request. He compares this behavior to an issue he had previously encountered with Opus 5.5. He also says it is new in his Fable testing: Fable 5 and Fable 5.1 had returned max-effort pelican images. The announcement does not give a price for the calls, change Anthropic access terms, or claim that the observed failure affects all Fable 5.2 users.
For existing llm users, version 0.31 means a newly released Anthropic model can be tested through an established CLI workflow. The announcement’s other useful detail is not a broad reliability verdict, but a visible failure mode: selecting the highest reasoning level can consume a large token budget without delivering the requested artifact.
What the archive adds
New Runtime covered the Fable 5.2 and Mythos 5.2 launch on November 9, 2026 and covered temporary Fable usage boosts for eligible paid subscribers on November 10, 2026. This is a new integration detail and a newly reported test outcome on the known Fable 5.2 story, not a restatement of the launch.
Trend signal
This supports the harness architecture outlives model choice pattern. A CLI harness can expose a new model through a familiar prompt workflow with a model-name change, preserving the surrounding tooling. The max run also illustrates the other half of the pattern: operational value depends on the harness’s handling of completion, token consumption, and failed outputs, not solely on which model it calls.
Why now
Willison released the plugin update after Anthropic’s November 9, 2026 release of Fable 5.2 and Mythos 5.2. The timing gives llm users a way to try Fable 5.2 immediately through the tool they already use, while the test also documents behavior observed during that first access window.
Why it matters
Model availability and useful task completion are separate engineering questions. Teams using llm can add Fable 5.2 to an existing command-line evaluation loop without rebuilding an integration, which lowers the cost of comparative testing. But Willison’s max-effort run shows how a workflow can spend substantial reasoning tokens and still return no requested artifact.
That matters most when a downstream step assumes output exists: an SVG renderer, parser, file upload, report generator, or automated reviewer may otherwise receive an empty result after waiting for an expensive call. Treat reasoning level as an evaluated configuration, not a universal quality dial. Compare xhigh and max on representative work, record token use and completion status, and decide which setting earns its additional latency and spend. The reported task is image generation, but the operational lesson applies to any agent or batch job whose success condition is a usable returned artifact.
Limits and uncertainty
The failed max run is one reported trial on one SVG prompt. It does not measure the failure rate across other tasks, repeated attempts, regions, API configurations, or clients. Willison does not provide a systematic quality evaluation of all five reasoning levels; his xhigh assessment is a visual judgement rather than a benchmark score. Anthropic has not, in this announcement, confirmed a model-wide defect or explained the observed behavior. The 128,000-token figure describes this failed run and is not a general cost estimate, a guaranteed ceiling, or a stated requirement for using Fable 5.2.
Skill to build now
Add an artifact-completion guard to one CLI model workflow this week. After a call, check that the expected file, valid JSON object, or other required output exists before handing the result to the next step. Log the model identifier, reasoning level, elapsed time, token usage when available, and completion state. If a high-effort call returns nothing, make the failure explicit and retry only under a tested fallback policy such as xhigh.
Reader takeaway
llm-anthropic 0.31 makes Claude Fable 5.2 selectable from the llm CLI, which is useful for fast evaluation. Its more consequential signal is narrower: one max-effort run spent 128,000 thinking tokens and returned no image, so validate completed outputs before treating higher effort as a better default.