Claude Opus 5 passed 24% of checkpoints on a rigorous new coding benchmark that tests whether AI models can maintain a codebase as requirements change over time, significantly outperforming Opus 4.8 and Sonnet 5 but still falling short of what production software engineering demands. The test, called SlopCodeBench, measures something most AI coding benchmarks ignore: whether a model can adapt when a specification evolves incrementally, the way real feature work actually arrives in software teams.

On July 25, 2026, researchers at humanlayer ran three of Anthropic's models through 17 sequential checkpoints across three coding problems of varying difficulty. The results paint a nuanced picture of progress that masks deeper fragility. Opus 5 wrote roughly 5 times more functions than Opus 4.8 across the same challenges, generating 29,065 lines of source code compared to around 9,000 for each of the other two models. But that volume advantage came with a catch: about half of Opus 5's code was tests rather than production logic, and the model still failed to complete any single problem cleanly.

What Makes SlopCodeBench Different From Other AI Coding Tests?

Traditional coding benchmarks like SWE-bench and Frontier Code show a model the entire problem specification at once and measure whether it can solve it correctly on the first try. SlopCodeBench does the opposite. It reveals requirements one checkpoint at a time, mimicking how software actually gets built in teams. A model might write code for a simple circuit validator in checkpoint one, then discover in checkpoint three that every signal needs to become a multi-bit vector with new operators and width checks. Code that assumed single-bit signals everywhere suddenly breaks.

This incremental structure exposes a critical failure mode that one-shot benchmarks miss: models that optimize for the immediate task without anticipating how the codebase needs to bend later. Scoring uses a strict pass rule: a checkpoint only counts if every new test passes and every regression test from earlier checkpoints still passes. Introduce a defect at checkpoint four, and checkpoints five through eight fail automatically, even if the model's code quality improves afterward.

Where Did Each Model Break Down?

The breakdown reveals where Opus 5 gained ground and where all three models hit a wall. Opus 5 passed 4 out of 17 checkpoints (24%), while both Opus 4.8 and Sonnet 5 passed only 1 out of 17 (6% each). Opus 5 stayed clean through the first three checkpoints of the circuit_eval problem, which was designed as the easiest of the three challenges. But then it introduced a defect at every single checkpoint afterward, a pattern that suggests strong early reasoning followed by compounding failures as complexity accumulated.

Opus 4.8 and Sonnet 5 both struggled earlier. Opus 4.8's only pass came on the first checkpoint of the database_migration problem, and its duplication metrics tell a revealing story: code duplication jumped from 4.6% to 16.8% between checkpoint one and checkpoint eight. That inflection point aligns exactly with checkpoint three of circuit_eval, where the single-bit-to-vector transition forced the model to copy-paste rather than refactor.

How to Interpret the Code Volume Difference

• Total Lines Written: Opus 5 generated 29,065 lines of source code versus roughly 9,000 each for Opus 4.8 and Sonnet 5, a roughly 3x difference in raw volume.
• Test Coverage Matters: About 51% of Opus 5's code was tests, compared to 24% for Sonnet 5 and 11% for Opus 4.8, meaning the production code gap was closer to 1.8x rather than 5x.
• Function Reuse Strategy: Opus 5 reused most of its roughly 2,000 functions across multiple places (only 14.9% were single-use), while Sonnet 5 wrote mostly one-off functions (71.5% single-use), suggesting different decomposition philosophies.

Whether Opus 5's extra test coverage actually bought correctness is ambiguous. Its early checkpoints held up better precisely where the other two models lacked coverage to catch regressions. But the strict-pass numbers suggest that extra testing was only partially effective, since Opus 5 still failed to complete any problem cleanly.

What Do These Results Mean for AI-Assisted Coding?

The SlopCodeBench results align with earlier findings that explainx.ai tracked across the Opus 5 launch cycle: today's frontier models still cannot be trusted to maintain a codebase without human oversight. Opus 5's 24% pass rate represents real progress over the 17% that Opus 4.6 achieved on the full benchmark, and it's consistent with incremental gains the model showed on other tests like ARC-AGI-3. But the pattern of strong early performance followed by compounding failures suggests the problem is not just raw capability but architectural: models struggle to reason about how code needs to evolve when requirements shift.

The honest assessment from the research is that every model's code got measurably worse as checkpoints piled up. The percentage of lines flagged as overly verbose rose from roughly 65% to 80% between checkpoint one and the final checkpoint. Complexity metrics diverged too: Opus 4.8's cyclomatic complexity (a measure of how many decision paths a function contains) climbed roughly 70% over the eight checkpoints of circuit_eval, while Opus 5 and Sonnet 5 kept complexity relatively flat. That suggests Opus 5 and Sonnet 5 found different ways to avoid the refactoring trap that caught Opus 4.8, but neither approach prevented eventual failure.

The benchmark itself remains unsaturated, meaning there is significant room for improvement before any model reaches saturation. The original SlopCodeBench paper reported GPT-5.4 at 11% strict pass rate, making Opus 5's 24% a meaningful step forward. But the small sample size (3 problems, one team, not independently replicated) means these results should be treated as a signal rather than a definitive verdict. The real takeaway is that SlopCodeBench has exposed a gap in how AI models handle the messy reality of software maintenance, and that gap remains wide even for the most capable models available today.