A new benchmarking update released in October 2026 evaluates how frontier AI models perform on SlopCodeBench, a test that measures coding agents' ability to maintain working code as requirements evolve. The dataset, published by the SlopCodeBench team, assesses models across 36 problems and 196 checkpoints, tracking both strict and core pass rates.

Strict pass rate, which counts only cases where a model passes every test including regression checks, ranged from 12.24% to 16.33% among tested models. GPT-6 Astra achieved the highest strict pass rate at 16.33%, narrowly outperforming GPT-6 SOL Medium at 15.31%. Opus 5.5 Medium recorded 14.29%, while Fable 5.1 XHigh registered 13.27%. Core pass rates, which are less stringent and include non-regression tests, varied from 65.82% to 73.98%, with Opus 5.5 Medium leading at 73.98%.

Cost analysis revealed that higher computational effort does not consistently improve outcomes. XHigh-tier processing, priced between 2.11x and 4.22x more than Medium-tier, increased strict pass rates only for GPT-6 Astra. For other models, strict pass rates declined or remained unchanged despite higher spending. GPT-6.1 Medium achieved the lowest estimated cost at $178.51, matching Astra Medium's strict pass count of 29.

Final checkpoint success did not always correlate with fully functional code. Only the last checkpoint of each problem was evaluated for strict pass status, and failed test cases were not treated as distinct bugs. Regression failures were documented separately. GPT-6 Astra XHigh had three regression failures, while Opus 5.5 Medium had none.

Code quality metrics indicated varying levels of technical debt. Opus XHigh produced the cleanest output, with 77.75% of its code triggering slop detection rules, compared to 94.25% for Opus Medium. The benchmark's scanner identified flagged lines related to repeated code and maintainability issues, with larger programs receiving greater weight in the assessment. Filtering out temporary files reduced Opus Medium's codebase by 41,265 lines, yet it still ranked highest in slop presence.

Model complexity scores showed GPT-6 SOL Medium had the highest average function and method complexity at 11.70, while Fable 5.1 recorded the lowest at 3.57. The report noted that splitting code into helper functions can reduce complexity without necessarily improving maintainability.

The findings highlight incremental progress in AI coding agents' ability to sustain code integrity through iterative development, though gains remain modest. The SlopCodeBench team plans to release a follow-up analysis examining qualitative patterns in model behavior.