SlopCodeBench
Explore score history ↗36 problems run as checkpoint sequences (196 checkpoints in v1.0): the agent extends its own earlier solution as new requirements arrive, so a checkpoint score reflects the accumulated codebase rather than a fresh attempt, and the headline erosion measures (verbosity, dead branches, redundant structure) are not a plain solve rate.
- Evaluates
- coding
- Openness
- unknown
- Importer
- Model reports
SlopCodeBench links
SlopCodeBench results and reported scores
Reported benchmark scores · Model reports
Reported scores: 19
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| Composer 2 | Cursor | 16.33 | 2026-04-24 | evidence |
| GLM 5.1 | Zhipu AI | 13.78 | 2026-04-24 | evidence |
| GLM 5.1 | Zhipu AI | 8.16 | 2026-04-24 | evidence |
| GPT 5.2-Codex | OpenAI | 21.94 | 2026-04-24 | evidence |
| GPT 5.3-Codex | OpenAI | 26.02 | 2026-04-24 | evidence |
| GPT 5.3-Codex-Spark | OpenAI | 8.16 | 2026-04-24 | evidence |
| GPT 5.4 | OpenAI | 23.47 | 2026-04-24 | evidence |
| GPT 5.4-Mini | OpenAI | 13.78 | 2026-04-24 | evidence |
| GPT 5.5 | OpenAI | 28.06 | 2026-04-24 | evidence |
| Kimi K2.5 | Moonshot AI | 7.14 | 2026-04-24 | evidence |
| Kimi K2.5 | Moonshot AI | 9.69 | 2026-04-24 | evidence |
| Kimi K2.5 | Moonshot AI | 8.67 | 2026-04-24 | evidence |
| Kimi K2.6 | Moonshot AI | 18.88 | 2026-04-24 | evidence |
| MiniMax M2.7 | MiniMax | 4.08 | 2026-04-24 | evidence |
| MiniMax M2.7 | MiniMax | 3.57 | 2026-04-24 | evidence |
| Opus 4.5 | Anthropic | 17.35 | 2026-04-24 | evidence |
| Opus 4.6 | Anthropic | 20.92 | 2026-04-24 | evidence |
| Opus 4.7 | Anthropic | 20.92 | 2026-04-24 | evidence |
| Sonnet 4.6 | Anthropic | 16.84 | 2026-04-24 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.
Which sources cite SlopCodeBench?
- SlopCodeBench leaderboard (v1.0, best version by % checkpoints view) SprocketLab, 2026-04-24
Independent audits of SlopCodeBench
- June Kim 2026-08-09
June Kim's audit reports that the erosion metric measures static code-shape drift, not yet extension robustness: its correlation with passing the next checkpoint is −0.018, weaker than raw line count at −0.212, and iteration is confounded because every checkpoint adds requirements. In his 12-problem sample, 28 of 343 core tests assert exact values no visible spec determines, and of all 196 answer keys run through the benchmark's own harness, 4 are deterministically defective and 1 is unstable.