Illustration, AI-generated from the story text. Tap a panel to read the story through it.

1 / 3Frequent compaction lets agents drop long context at little cost in accuracy, a new paper reports×

The finding

"You don't need long context," from Saurav Muralidharan and colleagues, finds that frequent compaction costs frontier models little on long tasks: accuracy moved between -3.7 and +5.6 points for Claude Opus 4.5 across six benchmarks.

The savings

Serving at 16,000 tokens instead of 256,000 cuts generation cost by 3.6 to 7 times on open models. On SWE-bench Verified, an agent limited to one action per context held 70.2% against a 70.0% baseline.

What it means

If this holds up, a lot of the memory now bought for long context buys less than planned.

Sources and the full story

‹›

Frequent compaction lets agents drop long context at little cost in accuracy, a new paper reports

"You don't need long context," from Saurav Muralidharan and colleagues, finds that frequent compaction costs frontier models little on long tasks: accuracy moved between -3.7 and +5.6 points for Claude Opus 4.5 across six benchmarks. Serving at 16,000 tokens instead of 256,000 cuts generation cost by 3.6 to 7 times on open models. On SWE-bench Verified, an agent limited to one action per context held 70.2% against a 70.0% baseline. If this holds up, a lot of the memory now bought for long context buys less than planned. SourcesA

Sources

• AYou don't need long context (arXiv)arxiv.org

This story ran in the Morning Brief of November 4th 2026.