ATLAS-Finance v1.1
Professional Work · 2026-10-15
Version 1.1.0 of Handshake AI's ATLAS-Finance: the same 100 investment-banking tasks across 13 simulated firms (inbox, chat, drive, calendar and notes over MCP), where agents must reconcile scattered, conflicting and mid-task-corrected inputs and deliver a live-formula Excel model. v1.1.0 regrades the trajectories with a revised verifier: rubric weights rescaled to a 1-10 integer scale, tolerances and gates moved to the section level, and criteria cleaned up; Handshake states that v1.1.0 scores are not comparable to v1.0.0. A rollout passes when 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 | 10.7 |
| Claude Fable 5.1 | 10.3 |
| GPT-6 Astra | 9.7 |
| Grok 4.6 | 4.7 |
| GPT-5.6 Sol | 4.0 |
| Gemini 3.8 Flash | 3.7 |
| Kimi K3 | 3.7 |
| Gemini 3.7 Flash | 3.3 |
| GLM-5.3 | 3.0 |
| Gemini 3.1 Pro Preview | 0.7 |