Claude Opus 5 Tops SlopCodeBench Eval

// 2h agoBENCHMARK RESULT

Claude Opus 5 Tops SlopCodeBench Eval

Humanlayer evaluated Claude Opus 5, Sonnet 5, and Opus 4.8 on a 17-checkpoint subset of SlopCodeBench, finding that Opus 5 led with a 24% strict pass rate compared to 6% for predecessor models. However, no model completed a multi-checkpoint problem without defects, and Opus 5 exhibited significant code inflation by writing five times as many functions as Opus 4.8.

// ANALYSIS

Current AI models remain incapable of running "lights-off" on complex software projects without active human steering and context management.

–**Evolving Requirement Signal:** SlopCodeBench addresses the key flaw of traditional single-shot coding benchmarks by testing how agents modify and maintain codebases across sequential checkpoints.
–**Opus 5 Marginally Leads:** Opus 5 achieved top performance among tested models (24% strict pass rate), but still failed to finish a single full problem without introducing bugs or breaking prior functionality.
–**Code Inflation and Slop:** High correctness in Opus 5 coincided with extreme verbosity (writing 29,000+ lines of code across the subset), proving that models still struggle with architectural clean-code discipline over time.

// TAGS

aicoding-agentllmbenchmarksslopcodebenchclaude-opus-5anthropicsoftware-engineeringai-codingagent

DISCOVERED

2h ago

2026-07-28

PUBLISHED

3h ago

2026-07-27

RELEVANCE

8/ 10

AUTHOR

dhorthy

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

RESEARCH37m ago

StateAct prioritizes program state over pixels for computer-use agents

StateAct is a research framework designed to improve long-horizon computer-use AI agents by tracking program state prior to processing visual pixel data. By maintaining structured state information and goal tracking throughout interaction steps, StateAct addresses the brittleness and reasoning decay associated with purely visual GUI automation, enabling agents to perform multi-step computer tasks with higher reliability.

INFRA1h ago

B.AI builds infrastructure layer for AI platforms

B.AI is shifting the focus of AI platform development away from weekly model releases toward building a comprehensive AI infrastructure. Rather than relying solely on incremental performance improvements, B.AI aims to provide the underlying platform capabilities and architecture necessary to power complex, next-generation AI applications.

OPEN SOURCE1h ago

Vercel Labs releases scriptc native TypeScript compiler

scriptc is an experimental compiler from Vercel Labs that transforms ordinary TypeScript code into lightweight, self-contained native executables without embedding V8 or Node.js. It lowers typed IR into LLVM or C while incorporating a QuickJS fallback for uncompiled npm dependencies.