ATLAS-Finance
Professional Work
Handshake AI's ATLAS-Finance places agents inside 13 simulated advisory firms with inboxes, chat, drives, calendars and notebooks served over MCP, and asks them to find an investment-banking assignment and its governing inputs, some contradicted or corrected mid-task, then build and deliver a live-formula Excel model. Its 100 tasks are graded by an agent-as-judge against expert rubrics of 42 to 486 weighted criteria. A rollout passes only if it meets every criterion weighted 3 or more and triggers no penalty; the headline is the pass rate over three rollouts per task.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 12.3 |
| Claude Fable 5.1 | 12.0 |
| Claude Opus 5.5 | 11.4 |
| GPT-6 Astra | 11.3 |
| Grok 4.7 | 5.0 |
| Gemini 3.8 Flash | 4.7 |
| Kimi K3 | 4.7 |
| GLM-5.3 | 4.7 |
| GPT-5.6 Sol | 4.7 |
| GPT-5.6 Luna | 0.3 |