Proximal Labs' ultra-long-horizon coding-agent benchmark: 17 open-ended technical projects spanning implementation, performance engineering, and applied ML research (e.g. optimizing a real compiler, inventing better ML optimizers, building a PostgreSQL-compatible server backed by SQLite). Agents get up to 20 hours per task and 5 trials each; tasks are graded 0–1 on partial progress, and frontier models barely make headway — making FrontierSWE one of the few unsaturated public coding benchmarks. Models are ranked by 'dominance' (win rate against a random opponent across tasks).
Real-world posts that speak to FrontierSWE — qualitative feedback from people using these models, each linked to its source. Never scored or merged with the leaderboard above.
Report tone
Report type
Topic
GLM 5.2 ranks unusually high on FrontierSWE (long-horizon agentic engineering) … using it with OpenCode is also not far from the quality of Claude Code or Codex.
Task Day-to-day agentic coding with GLM-5.2 in OpenCode.