{"slug": "nvidia-study-finds-ai-agents-keep-accuracy-with-short-contexts", "title": "Nvidia study finds AI agents keep accuracy with short contexts", "summary": "An Nvidia paper titled \"You don't need long context,\" authored by Peter Belcak, Pavlo Molchanov and nine co-authors, found that AI agents can match full-history accuracy by compacting their conversation history into short handoff notes, shrinking working context to as little as 16,000 tokens. On SWE-bench Verified, Anthropic's Claude Opus 4.5 scored 70.2% with one action per context versus 70.0% with full history, and tuning compaction settings lifted the score to 74.8%; across configurations accuracy shifted from −3.7 to +5.6 points for Opus 4.5 and −3.3 to +4.4 points for OpenAI's GPT 5.6 Sol across thirteen benchmarks, six of them long-horizon tasks including Terminal-Bench 2.1. Peak context in single-action tests fell from 130K–200K tokens to 15K–37K, with median context at 7K–9K tokens, and the researchers report long-context inference can cost up to 7x more at 256K tokens than at 16K.", "body_md": "# Nvidia study finds AI agents keep accuracy with short contexts\n\nA new paper suggests agents that compress their history into handoff notes can match full-history performance at a fraction of the context\n\nThe AI industry has spent a lot of energy building models that can hold more and more text in memory at once. A new [Nvidia](https://cryptobriefing.com/markets/nvidia/) paper suggests agents may not need most of it.\n\nThe research, titled **“You don’t need long context,”** found that AI agents can preserve accuracy on complex tasks by regularly compacting their conversation history into short handoff notes. Researchers were able to shrink the working context to as little as 16,000 tokens while still matching full-history performance.\n\n## What the researchers actually tested\n\nThe paper is authored by Peter Belcak, Pavlo Molchanov, and nine co-authors. Its focus is frontier agents, meaning AI systems built to work through long, multi-step tasks rather than answer a single question.\n\nInstead of dragging the entire conversation forward at every step, the agent periodically condenses what has happened into a handoff note. The next step starts fresh with only that summary.\n\nThe team ran the approach on two models: [Anthropic](https://cryptobriefing.com/markets/anthropic/)’s Claude Opus 4.5 and [OpenAI](https://cryptobriefing.com/markets/openai/)’s GPT 5.6 Sol. Testing covered thirteen benchmarks. Six of those were long-horizon tasks, including SWE-bench Verified and Terminal-Bench 2.1.\n\n## The numbers behind the claim\n\nThe headline result came from Claude Opus 4.5 on SWE-bench Verified, a software engineering benchmark. With one action per context, where the agent hands off after every single step, the model scored 70.2%.\n\n### AI, tech, and the markets they move—in one daily briefing.\n\nDaily. Free. Join 34,000+ readers across crypto, finance, and policy.\n\nWith its full history available, it scored 70.0%. So the agent that kept forgetting everything except its notes did marginally better than the one that remembered it all.\n\nThe researchers also found that tuning the compaction settings could lift Opus 4.5’s score on that task as high as 74.8%.\n\nAcross configurations, the accuracy changes were generally modest. For Opus 4.5, shifts ranged from −3.7 to +5.6 points against the full-history baseline. For GPT 5.6 Sol, the range ran from −3.3 to +4.4 points.\n\nIn the single-action compaction tests, peak context dropped from 130K–200K tokens to between 15K–37K. Median context sat between 7K–9K tokens.\n\n## Why the context window race may be the wrong race\n\nNvidia’s researchers argue the better lever is how context gets managed. The paper points to efficient summarization, selective reading of information, and structured handoffs as more effective ways to handle agentic workloads than simply stuffing more tokens in.\n\n## What this means for AI builders and their budgets\n\nThe most immediate implication is cost. According to the research, long-context inference can run up to 7x more expensive for models using 256K tokens compared with those using 16K tokens.\n\n**Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).", "url": "https://wpnews.pro/news/nvidia-study-finds-ai-agents-keep-accuracy-with-short-contexts", "canonical_source": "https://cryptobriefing.com/nvidia-study-ai-agents-short-contexts/", "published_at": "2026-11-04 18:25:29+00:00", "updated_at": "2026-11-04 18:52:05.222651+00:00", "lang": "en", "topics": ["ai-agents", "large-language-models", "ai-research", "artificial-intelligence"], "entities": ["Nvidia", "Peter Belcak", "Pavlo Molchanov", "Anthropic", "Claude Opus 4.5", "OpenAI", "GPT 5.6 Sol", "SWE-bench Verified"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/nvidia-study-finds-ai-agents-keep-accuracy-with-short-contexts", "markdown": "https://wpnews.pro/news/nvidia-study-finds-ai-agents-keep-accuracy-with-short-contexts.md", "text": "https://wpnews.pro/news/nvidia-study-finds-ai-agents-keep-accuracy-with-short-contexts.txt", "jsonld": "https://wpnews.pro/news/nvidia-study-finds-ai-agents-keep-accuracy-with-short-contexts.jsonld"}}