Sources
A 17-Model, 28-Task Single-Harness Leaderboard Puts GLM-5.3 at 100% Pass for $0.28 a Lap, About a Fifth of GPT-5.5
Ed Yau's Ed-o-meter, updated 2026-08-23 from a source run stamped 20260822T172041Z, runs seventeen models through the same 28 real-world tasks with verbatim prompts and deterministic grading, and publishes pass rate, rubric score, security, TTFT, and cost per task side by side. GLM-5.3 takes the top slot at 100% pass and a 9.3 rubric for $0.28 per lap; GPT-5.5 is faster at 13.2s TTFT versus 16.3s but costs roughly five times more. Read the caveats before quoting it: single trial per task, wide Wilson intervals, and the rubric scores were assigned retroactively by Fable 5 against saved answer text.
↳ Follow the thread