Blog
Notes from the Benchmark
Model audits, benchmark design, and why we measure what we measure.
Does the Harness Matter? Lessons from ALE-Claw on Agents’ Last Exam
Swapping the model under a fixed harness moves ALE pass rate by 18 points; swapping the harness moves it by 5–6. ALE-Claw, a deliberately minimal harness, matches heavier ones at a fraction of the cost. Interactive: pick any agents to compare cost, time, and tool-call mix.
Read more →Codex (GPT-5.5) vs. Claude Code (Fable 5) vs. Cursor (Composer 2.5): An Interactive Comparison
A live, interactive comparison of agentic coding harnesses on the agenthle benchmark. Toggle any model to compare performance, cost and time, per-domain strengths, and a unified failure-mode taxonomy built by analyzing run trajectories.
Read more →Agents’ Last Exam
We evaluated Fable 5, GPT-5.5, Composer 2.5, and other frontier systems on 1,500+ expert-sourced tasks spanning 55 occupations. Today’s agents solve a meaningful fraction of professional work, but on ALE’s hardest tier every frontier agent we tested, including Fable 5, scored 0%.
Read more →The Benchmarks Keep Falling. The Economy Hasn’t Noticed.
AI progress is shaped by what we choose to measure. We are building the instrument that measures what matters — real economic value, not proxy metrics.
Read more →