The efficiency gains are huge. On #SWEbench Multilingual, it boosted success by 5 percentage points and cut execution costs by 38.5%. Plus, with only $4.03 in search costs, it created an agent that solved 82.0% of #TerminalBench 2.1 tasks, matching top scorers like #Codex!

A verified version of SWE-Bench Pro addresses reward hacking via leaked gold solutions and task quality issues like misleading problem statements, with anti-hacking safeguards and minimal task refinements.

#SWEbench #AIAgents #BenchmarkReliability
https://arxiv.org/abs/2609.08149

Claude Opus 5.5 hit 89.9% on SWE-bench Pro multi-file tasks. We updated our model-proxy to keep Anthropic SSE streams alive during extended thinking phases, preventing…

https://monkeyscode.com/capuchin?utm_source=bluesky&utm_medium=company&utm_campaign=story-102ae2e2

#SWEbench #Anthropic

AgentEvolver adds a system-wide self-evolution loop around task execution, exposing eight entity families to a shared versioned lifecycle while keeping the base model fixed. It reports 82.08% on SWE-bench Pro Public,…

#AgentEvolver #AIInfra #AgenticAI #SWEbench
https://arxiv.org/abs/2610.11613