Claude Fable 5 Refuses Mid-Tool-Call on Mundane Tasks, Risking Agent State Corruption

Anthropic's Claude Fable 5 model triggered mid-tool-call refusals on 11 of 44 turns during coding-agent evaluations conducted on July 5, 2026, even on trivial tasks like fixing a config default or booking a meeting room. The refusal arrives via a stop_reason field set to 'refusal' partway through generating tool arguments, yet the truncated output still parses as valid JSON, giving agent loops no obvious signal to halt execution. Any agent that runs tool calls without checking the stop reason will write incomplete files to disk, corrupting workspace state with no recovery path within the same conversation context. Cost comparisons showed Fable 5 running roughly 15 times more expensive than GLM-5.2 on a four-turn coding task, though only about 5 times pricier than claude-sonnet-5 on batch workloads. Competing models — claude-opus-4-8, claude-sonnet-5, and glm-5.2 — each achieved 100% pass rates on the same trivial tasks where Fable 5 passed only 58–75% of episodes, with every failure traced to a refusal rather than incorrect code.