Alibaba's OpenGCUA Lifts 8 Models by 19.6 Points on OSWorld-v2
TL;DR
- The same harness lifts all eight frontier models uniformly, from cheapest to most capable, confirming scaffolding rather than model weights drives the performance gap.
- Steps per task drop 53% alongside the +19.6 accuracy gain, making OpenGCUA a cost lever and a quality lever simultaneously.
- GPT-6 Astra's step count falls from 102 to 34 per task under OpenGCUA, a concrete efficiency signal for enterprise buyers weighing API costs at $10 per million input tokens.
Across eight frontier models, Alibaba's Tongyi-MAI team reports a +19.6-point mean partial-score lift on OSWorld-v2 when each one runs inside its new OpenGCUA stack, with reported steps per task falling 53.0% versus a batched GUI-only baseline. The paper says the gap between the strongest and weakest evaluated model 'shrinks by more than half.'
OpenGCUA is what the authors call a generalist computer-use agent, defined in the paper as 'a system in which the model chooses how to observe, act, and organize its work across visual and programmatic interfaces.' GUI clicks, accessibility calls, API calls, and shell commands all live inside one program with loops, branches, and intermediate results. The second design move is 'moving selected observations and task-state decisions from fixed harness rules to the model,' which the paper labels harness-to-model decision transfer.
The evaluated models include GPT-6 Astra, Claude Opus 5, and Qwen3.8-Max.
There is a cost side. 'Selected Qwen3.8-Max runs use 19.4 times as many reported reasoning tokens as GPT-6 Astra,' the paper notes, on a set of four real macOS tasks. The sample is too small to generalize, but Tongyi-MAI published the figure anyway.
The release bundles a macOS menu-bar harness for real-device execution, a cross-device SDK spanning computer, browser, and phone, and the full evaluation trajectories from all eight models. It is the third Tongyi/Alibaba write-up on our tracker this week, after Waypoint-OPD on distillation gating.
What others are reporting
-
Tongyi-MAI (GitHub) Read →
First-party README documents the three-axis decision transfer from harness to model across observation, action selection, and task organization, with GPT-6 Astra step count dropping from 102 to 34.
GPT-6 made computer use the frontier. We make it open.
-
AICoder Read →
Developer-facing framing with live leaderboard links and cost context; anchors GPT-6 Astra at $10/M input tokens, converting the 53% step reduction into a concrete API-cost argument.
OpenGCUA elevates scores on OSWorld-v2 across all eight tested models...by +19.6 points while cutting steps by 53.0%.
-
CCTest Read →
Systems-level framing focused on design philosophy: the harness hands observation, action-mode selection, and task-tracking decisions to the model rather than enforcing fixed rules.
The agent can also read the accessibility tree, call application or system APIs, execute shell commands, and return to visual interaction when structured information is insufficient.
Originally reported by huggingface.co
Read the original article →Original headline: Alibaba Opens OpenGCUA, a Full Stack for Generalist Computer-Use Agents That Lifts 8 Frontier Models by 19.6 Points on OSWorld-v2