Atlas

Benchmarks

← All benchmarks

BizAgentBench (strict success)

Agents · 2026-10-14

50 long-horizon, multi-turn payment-operations tasks built from Ant International's real procedures, in a shared environment of 789 knowledge documents and 601 discoverable business tools with an LLM customer simulator. Tasks link workflows through shared resources, timing and prior commitments. Strict success requires the customer goal and every process and protection requirement (approvals, funds and policy in force at each action), in percent.

Top models (higher is better)

ModelScore
Claude Opus 564.0
GPT-6 Sol61.3
DeepSeek V4.1 Flash43.3
GLM-5.3 Flash43.3
Muse Spark 1.338.0
Qwen3.8 27B32.7
Qwen3.5 397B A17B0.7
Qwen3.5-9B0.0
Loading Atlas data…