GLM 5.3 SlopCodeBench Results

Howdy once again, I had a request to try out 5.3 on the benchmarks - they're unsaturated so it's a fun test right now! This one was interesting because I accidentally ran it on all 36 problems (rip $200) instead of the 9 i typically do previous runs a b c benchmark context: the ai is tasked to build a tool step by step, we add new requirements mid way - it has to handle new things without breaking the old things - the benchmark "hides" the old tests So it didn't magically solve the benchmark just like all the other AIs, i don't think there has been a single successful solve yet, but it did tie with Fable/Sol in the same subset of problems! Another interesting thing that we expect is that the more difficult the problem, the more $$ it gets - so difficulty seems to be correlated with token output in order to solve them AI;DR full results On the three-problem, 17-checkpoint list from the Opus 5 report, GLM scored 8/17 strict (47.1%). On the six-problem, 30-checkpoint list from the Fable, Sol, and Kimi report, it scored 10/30 (33.3%), tied at the strict threshold with Fable 5 and GPT-5.6 Sol. Reported system Opus list · 17 Fable/Sol/Kimi list · 30 GLM-5.3 · pi 8 (47.1%) 10 (33.3%) DeepSeek V4 Flash 0731 · pi B 5 (29.4%) — Opus 5 · Claude Code 4 (23.5%) — Qwen3.8-27B · pi 3 (17.6%) 4 (13.3%) DeepSeek V4 Flash · OpenCode 3 (17.6%) — Fable 5 · Claude Code — 10 (33.3%) GPT-5.6 Sol · Codex — 10 (33.3%) Kimi K3 · Modal / OpenCode — 8 (26.7%) Kimi K3 · Baseten / OpenCode — 7 (23.3%)