Working at the frontier: How Cursor knew Claude Fable 5 was ready for the hardest 1% of problems
02:00 · July 17, 2026 · Claude Blog
Summary
Cursor, the AI coding environment that integrates multiple frontier models, found that public benchmarks increasingly diverged from how developers actually use these systems on ambiguous tasks. To address the gap, engineer Nate Schmidt’s team created CursorBench, an evaluation suite built around underspecified prompts that mirror real workflows: a stack trace accompanied only by the word “fix,” or instructions that deliberately point to the wrong module. The benchmark measures whether a model can infer intent, locate root causes, validate changes, and report results without further scaffolding.
On this suite Claude Fable 5 reached 72.9 percent at maximum effort, the highest score recorded. Traces from the hardest items showed the model performing global reasoning—considering system-wide constraints and long-term consequences—rather than the local, step-by-step adjustments typical of earlier models. It also completed the same tasks with fewer tokens, suggesting more efficient internal planning. In one internal test, the model was given a blank prompt to land a simulated spacecraft on the moon; within hours it executed an orbital reconnaissance flight to gather telemetry before attempting the final descent, whereas prior models exhausted resources without ever reaching the surface.
Cursor’s engineers now route the hardest problems—large-scale refactors, nuanced edge-case analysis, and previously deferred architectural work—to Claude Fable 5 while delegating routine edits to lighter models. The division reduces context-switching overhead and lowers the activation energy required to tackle tasks whose solution path is not yet clear. The same capability supports lightweight coordination inside the team: an agent can review a colleague’s recent commits and surface potential conflicts before either developer interrupts their own work.
Schmidt continues to test the model’s limits on longer-running, unattended backend systems and is developing more realistic evaluation environments that capture multi-day agent behavior. The practical outcome is that certain classes of previously intractable problems now appear worth attempting.
Why it matters
This article provides actionable insights for product teams and builders on evaluating and deploying advanced AI models for software engineering. It highlights practical strategies like model routing and custom benchmarking that Dutch AI practitioners can implement to optimize development workflows.
