{
 "id": "surf-data-platform-onchain-social-prediction-crypto-ai-agent-intelligence/surf2-f2",
 "title": "AI model tool selection",
 "subtitle": "General AI models route crypto queries to web search instead of onchain data, and accuracy suffers.",
 "note": "Example query: \"Is contract 0x742d…f2bF3e8 safe to interact with?\" Web search returns press coverage and misses a phishing drainer; a block explorer flags the deployer and blocks the interaction. Surf co-authored CAIA through Cybertino Lab.",
 "source": "Source: CAIA benchmark (17 models, 178 tasks), Vals.ai Finance Agent Benchmark (Apr 2025).",
 "sources": [
  {
   "name": "CAIA benchmark",
   "url": "https://arxiv.org/abs/2510.00332"
  },
  {
   "name": "Vals.ai Finance Agent Benchmark",
   "url": "https://www.vals.ai/benchmarks/finance_agent"
  }
 ],
 "report": {
  "title": "Surf: A Data Platform for Crypto Intelligence",
  "url": "https://www.khala.io/surf-data-platform-onchain-social-prediction-crypto-ai-agent-intelligence",
  "date_published": "2026-04-03"
 },
 "url": "https://www.khala.io/surf-data-platform-onchain-social-prediction-crypto-ai-agent-intelligence#ix-surf2-f2-t",
 "publisher": "Khala Research",
 "cite": "Khala Research, \"AI model tool selection\", in \"Surf: A Data Platform for Crypto Intelligence\", April 3, 2026. https://www.khala.io/surf-data-platform-onchain-social-prediction-crypto-ai-agent-intelligence#ix-surf2-f2-t",
 "csv": "https://www.khala.io/data/surf-data-platform-onchain-social-prediction-crypto-ai-agent-intelligence/surf2-f2.csv",
 "columns": [
  "Label",
  "Value",
  "Detail"
 ],
 "rows": [
  {
   "Label": "Web search chosen",
   "Value": "55.5%",
   "Detail": "Share of tool selections that pick web search when specialist tools are available."
  },
  {
   "Label": "No tools",
   "Value": "12% to 28%",
   "Detail": "Frontier model accuracy without tools."
  },
  {
   "Label": "Best model with tools",
   "Value": "67.4%",
   "Detail": "Best general model, with the correct tools available."
  },
  {
   "Label": "Human analyst",
   "Value": "80%",
   "Detail": "Baseline on the same task set, 12.6 points above the best model."
  },
  {
   "Label": "Finance tasks",
   "Value": "48.3%",
   "Detail": "Best general model (OpenAI o3) on Vals.ai Finance Agent; none passed 56%."
  },
  {
   "Label": "Surf vs no-tool models",
   "Value": "4x",
   "Detail": "CAIA outperformance; about 1.2x against the best tool-augmented model."
  }
 ],
 "quantitative": true
}