SlopCodeBench 10/2026 – Opus 5.5, Astra wins overall, xhigh considered harmful
Higher effort cost 2.11–4.22 times as much, but improved strict results in only one of three model pairs.
Astra XHigh led with 32 of 196 checkpoints passing every test category, including regression tests. Opus 5.5 and GPT-6 SOL both passed fewer checkpoints at XHigh than at Medium effort. Each setting ran once, so the results cannot establish that extra reasoning caused the declines. Even the best campaigns finished with fully passing code on only three of 36 problems.
HN · GitHub AI's note
Astra XHigh led with 32 of 196 checkpoints passing every test category, including regression tests. Opus 5.5 and GPT-6 SOL both passed fewer checkpoints at XHigh than at Medium effort. Each setting ran once, so the results cannot establish that extra reasoning caused the declines. Even the best campaigns finished with fully passing code on only three of 36 problems.
HN · GitHub AI's note
score 6
Discussions
- hn · 129 points · 55 comments
- hn · 142 points · 63 comments
- hn · 151 points · 67 comments
- hn · 160 points · 72 comments
- hn · 165 points · 74 comments
- hn · 177 points · 80 comments
- hn · 188 points · 81 comments
- hn · 194 points · 84 comments
- hn · 202 points · 86 comments
- hn · 209 points · 86 comments
- hn · 218 points · 88 comments
- hn · 228 points · 88 comments
- hn · 234 points · 88 comments
- hn · 239 points · 91 comments
- hn · 245 points · 93 comments
- hn · 247 points · 94 comments
- hn · 248 points · 95 comments
- hn · 253 points · 96 comments
- hn · 258 points · 100 comments