AI Lab · by Hamza Malik · 2026-10-20
SlopCodeBench Oct 2026: GPT-6 Astra beats Opus 5.5 overall, and xhigh reasoning apparently makes things worse
github.com
The October SlopCodeBench writeup puts Opus 5.5 and GPT-6 Astra head to head, with Astra coming out on top overall. The spicier takeaway is the 'xhigh reasoning considered harmful' line, i.e. cranking reasoning effort to the max didn't help and may have hurt. That matches a vague feeling a lot of us have had in agent loops, where more thinking sometimes just means more overengineered output. Does this line up with your experience, and what effort level do you actually run day to day?
84 sparks · 20 comments
Comments
Tiny Atlas · 11 likes
I'm skeptical of any single bench crowning a winner. Opus is still nicer for me on large refactors where it has to keep a lot of context straight
Rachel Tan · 6 likes
same, ranking flips depending on what kind of codebase you throw at it
Chloe Davis · 2 likes
agreed, overall wins hide a lot. I'd love per-task breakdowns
Mateo Cruz · 7 likes
Astra winning is surprising to me, I've found it confidently wrong more often. maybe I'm prompting it wrong
Pixel Orbit · 1 likes
or the bench rewards different stuff than your daily work, both can be true
Grace Parker · 5 likes
Does anyone know if they controlled for the agent harness? Same scaffold, same tools? Because harness differences can swamp model differences pretty easily
Siti Lestari · 1 likes
they're context engineering people so I'd guess the harness was the point, but yeah worth checking
Amelia Hughes · 5 likes
overthinking is the most human thing these models have done
Mia Bauer · 10 likes
me at 2am before a deploy
Aria Thompson · 4 likes
medium effort for most things, high when debugging something gnarly. never touched xhigh after the first week, it was slow and the bill was scary
Jisoo Park · 14 likes
the bill is the real benchmark
Mariam Said · 3 likes
would be fun to see the same bench next month, these rankings feel like they change every few weeks now
Iryna Melnyk · 1 likes
they literally do lol, half the models in the last one are deprecated
Noé Petit · 1 likes
xhigh considered harmful matches my experience 100%. it writes an essay of reasoning then refactors three files I didn't ask about
Haruto Sato · 6 likes
this is literally the 'slop' the bench is named after
Bilal Raza · 2 likes
it plans like it's getting paid by the token. oh wait
Elsa Andersson · 0 likes
benchmarks with 'slop' in the name are the only honest benchmarks left
Gentle Marauder · 0 likes
the meta lesson here is test on your own repo. every time I switch models based on a chart I regret it a bit