Agentic tool use

A model's score is built only from results that someone other than its maker ran.

• The scoring set is the benchmarks chosen for this specialty. A benchmark takes part only when at least 5 models have an independent or third-party result on it.
• Each result is rescaled to 0 to 1, and a model's best setting counts.
• The score is the average over the benchmarks a model has, times 100.
• A model ranks only when it has results on enough of them. The rest show "Not enough independent results yet".
• Self-reported results are shown but never counted.

Method: category-2026-09-29.1

Scoring set: AutomationBench (and AutomationBench-AA), EnterpriseOps-Gym, MCP-Atlas, τ³-bench (tau3-bench), Toolathlon (Tool Decathlon), BALROG, Deep Research Bench (FutureSearch).

Effective set (used for the score today): BALROG, Deep Research Bench (FutureSearch)

• AutomationBench (and AutomationBench-AA): no importable independent results yet
• EnterpriseOps-Gym: no importable independent results yet
• MCP-Atlas: no importable independent results yet
• τ³-bench (tau3-bench): no importable independent results yet
• Toolathlon (Tool Decathlon): no importable independent results yet

• AutomationBench (and AutomationBench-AA)ActiveHeadroomCounts toward the score0 results
• BALROGActiveHeadroomCounts toward the score41 results
• BrowseCompActiveSaturated0 results
• Deep Research Bench (FutureSearch)ActiveHeadroomCounts toward the score41 results
• EnterpriseOps-GymActiveHeadroomCounts toward the score0 results
• MCP-AtlasActiveNear-ceilingCounts toward the score0 results
• Toolathlon (Tool Decathlon)ActiveHeadroomCounts toward the score0 results
• Vending-Bench 2ActiveHeadroom0 results
• τ²-bench (tau2-bench)SupersededSaturated0 results
• τ³-bench (tau3-bench)ActiveHeadroomCounts toward the score0 results