Simon Willison's first 5 hours with Claude Fable 5

Simon Willison spent the first 5.5 hours with Claude Fable 5 running real projects, and his writeup is the most useful independent read on the model so far. The headline number on price is that Fable lists at $10 per million input tokens and $50 per million output, roughly double Claude Opus 4.8, and on subscription plans it is available only until June 22 before pricing settles. Willison spent $110.42 total over the session, with $99.26 of it on a single project.

The two work items are the more interesting evidence. He pointed Fable at his own micropython-wasm project and asked it to upgrade from MicroPython to full CPython compiled to WebAssembly. Fable handled most of the architectural work on its own and produced a working 13.9 MB wheel. He then asked it to add a human-in-the-loop approval mechanism to datasette-agent, which required reaching into and refactoring the underlying llm Python library; Fable did that too, and the result shipped as llm 0.32a3 with new tool-call pause-and-resume primitives. A separate knowledge test, asking the model to list his own open-source projects, returned a noticeably wider list than Opus 4.8, suggesting Fable has more parametric knowledge baked in.

Why it matters

If you are deciding whether to budget for Fable, the price gap is the question: at 2x Opus, the test is whether the capability difference on your specific workloads pays for itself, and Willison's tasks are a useful benchmark to compare yours against. The June 22 cutoff on subscription plans also means anyone running Fable in production should plan for a pricing change before that date.