Claude Fable 5.1 reaches 5/5 at xhigh
Claude Fable 5.1 rose from 1/5 to 5/5 on Terminal-Bench 3.0’s HTML sanitizer task when effort increased from low to xhigh. The result shows extra test-time compute can improve security verification, but with substantially higher latency and token costs. [Source](https://claude.dev/blog/spending-your-effort/)
This is a strong argument for routing effort by risk rather than using xhigh by default. The result demonstrates that Fable 5.1 can verify aggressively when given enough compute, not that it writes secure code flawlessly on the first pass.
- –Low effort produced a quick one-pass filter tested against a single page; higher effort inspected parser behavior, ran XSS suites, and built a fuzzer.
- –Fable 5.1’s Terminal-Bench security pass rate improved from 64% at low effort to 87% at its highest settings.
- –More effort cannot fix a fundamentally wrong interpretation; max-effort runs still showed more wrong-reading failures than low-effort runs.
- –A practical workflow is low effort for implementation, high effort for review, and xhigh or max for security-critical or edge-case-heavy work.
DISCOVERED
1h ago
2026-09-26
PUBLISHED
1h ago
2026-09-26
RELEVANCE
AUTHOR
DIY Smart Code