Code modification 70.2%… Claude Opus 4.5, record compression test

NVIDIA researchers reported that Claude Opus 4.5 solved 70.2% of code-repair tasks when its action history was summarized after each step, versus 70.0% when the full uncompressed history was retained, according to a technical report titled “You don’t need long context” posted to arXiv on the 4th (Korea time). The study compared passing compressed conversation histories as fresh context against keeping complete records, testing Claude Opus 4.5 and GPT 5.6 Sol across six long-horizon tasks and extended workflows. Context in the paper refers to the instructions, prior dialogue, and tool outputs a model reads in a single call, which accumulate as an agent works through multi-step tasks and raise input loads on subsequent calls. The near-identical scores suggest history compression can match full-context performance while cutting the input burden of long agentic runs. What to watch next is whether these results hold across a broader task set and other models, and whether compression becomes standard practice for agent deployments.