# OpenAI Pushes GPT-6 Astra Into Enterprise Workspaces, Led by Computer-Use Scores

The model reports 72.6 percent on OSWorld 2.0 and more than doubles its predecessor on AutomationBench, with all published figures coming from OpenAI's own evaluations.

- Published: 2026-09-10T06:24:02.019Z
- Canonical: https://polylog.news/ai/2026-09-10/openai-pushes-gpt-6-astra-into-enterprise-workspaces-led-by
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [OpenAI News](https://openai.com/index/gpt-6-astra-next-generation-work), [Polylog editors](https://polylog.news), [arXiv stat.ML](https://arxiv.org/abs/2609.09245)

OpenAI followed the September 3 launch of GPT-6 Astra with a [business-focused rollout](https://openai.com/index/gpt-6-astra-next-generation-work), positioning the model for work through reasoning, computer use, and writing and design judgment. Access reaches Pro, Business and Enterprise plans and the application programming interface (API), with [enterprise administrators required to switch it on](https://fortune.com/2026/09/03/openai-debuts-gpt-6-astra-computer-use-greg-brockman-says-start-of-agi/) because it is off by default in managed workspaces.

The published figures focus on the model's ability to operate software. Astra reports 72.6 percent on OSWorld 2.0 with roughly 47 percent less time per task than GPT-5.6 Sol, and AutomationBench rises from 18.1 percent to 41.4 percent, indicating that longer sequences of actions complete without human intervention. On security capability, OpenAI reports 100 percent on ExploitBench against 78.5 percent for the prior model, and 42.4 percent on ExploitGym against 30.3 percent. API pricing is 10 dollars per million input tokens and 50 dollars per million output tokens.

Every one of those figures is vendor-reported, and a more accurate reading is narrower than the marketing suggests. Independent write-ups conclude Astra is a clear efficiency and computer-use gain over OpenAI's own prior models rather than a decisive move past rival models. A paper posted to arXiv this week adds a specific caution: [fixed-rollout pass@k evaluations](https://arxiv.org/abs/2609.09245), a method that tests success across only a fixed number of attempts per task, capture just a limited set of quantities from the samples actually collected, so extrapolating success rates far beyond that sampled budget is not supported by the data.

Demonstrations are circulating faster than measurements. One widely shared clip shows Astra [directing a robot arm holding a paintbrush](https://t.me/aipost/8096) and improving across attempts, an impressive demonstration that lacks a controlled baseline for comparison.

## What this means

Computer-use scores are now the dimension OpenAI is competing on, which shifts enterprise buying from chat quality to how many steps of internal software a model can complete unattended. Software vendors whose value sits in the user interface lose leverage when an agent operates the interface instead of a person, while OpenAI gains a reason to charge work-tier prices. The unresolved question is whether independent evaluations on real enterprise environments reproduce the OSWorld gap or show it narrowing under messier conditions, and enterprise pilots will settle that within a quarter.

## What to watch

- Independent reproductions of the OSWorld 2.0 and AutomationBench figures on tasks OpenAI did not select.
- How many enterprise administrators actually enable Astra, since default-off deployment is a real brake on adoption numbers.
- Whether OpenAI publishes per-task cost for computer use, which decides if agentic workflows are cheaper than the labor they replace.