OpenAI today unveiled GPT-6 Astra, a model engineered for professional and enterprise workflows that the company describes as its most capable and aligned model to date. GPT-6 Astra is designed to reduce the integration and engineering overhead companies typically face when adopting AI by interacting directly with desktop applications, applying corporate templates and automating complex technical tasks without requiring bespoke connectors.
Astra’s desktop and browser capabilities are a central focus. In ChatGPT Work and Codex, Astra can operate everyday desktop software even when those applications lack a native API, enabling teams to apply the model to existing processes without redesigning systems. OpenAI highlighted internal and customer use cases to illustrate practical benefits: in one benchmarking demonstration Astra completed tasks drawn from the Microsoft Excel World Championship roughly four times faster than the human winner, allowing analysts to spend more time interpreting results rather than building spreadsheets.
The company also described technical uses inside its teams. Engineers used Astra and Codex to produce a multicamera developer impressions video and to diagnose a memory-allocation bottleneck in Codex sessions. Adjusting allocators in that test environment produced a 25× reduction in turn latency while increasing peak memory usage by about 30%. OpenAI says Astra more consistently matches company voice, templates and design standards, producing outputs that need less post-editing to meet corporate formatting and tone.
OpenAI positioned cost-efficiency as another pillar of Astra’s design. The model was trained to accomplish tasks with fewer tokens and fewer retries, which OpenAI says reduces rework and lowers per-task cost. Published starting prices for Astra are $10 per million input tokens and $50 per million output tokens. On industry evaluations, Astra earned 57.9% on Terminal-Bench 4.0, a suite for complex terminal-based tasks such as engineering, system configuration and data analysis—compared with 37.3% for GPT-5.6 Sol and 55.8% for Anthropic’s Claude Fable 5.1. OpenAI reported approximately 9% and 63% lower estimated API cost per task compared with GPT-5.6 Sol and Claude Fable 5.1, respectively.
Safety and administrative controls are emphasized for consequential work. OpenAI reports that Astra produced unintended outcomes far less often than prior models and competitors on an internal computer-use safety benchmark that simulates high-risk business scenarios, such as accidental disclosure of confidential data or destructive changes. Specifically, Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1 in that evaluation. Astra is also the first model OpenAI says has met the Critical cybersecurity capability threshold under its Preparedness Framework.
To support enterprise deployment, OpenAI added administrative controls that let organizations restrict which websites and desktop applications Astra can access, manage file uploads and downloads, and control browsing history. ChatGPT Work and Codex include confirmation policies that can require human approval before the model executes consequential actions, alongside automated review of potentially unsafe or unauthorized tool calls. These features are intended to let teams begin with a constrained configuration and expand access as confidence grows. OpenAI also offers Zero Data Retention for eligible API customers on supported endpoints, subject to approval.
Astra is broadly available through ChatGPT Work, Codex and the API. OpenAI announced new enterprise plugins for ChatGPT Desktop that leverage Astra’s browser capabilities, including integrations with Oracle Analytics, Power BI (a Microsoft Fabric service), Navan and Avalara. These plugins are described as a way to more easily connect Astra to familiar enterprise systems and analytic tools.
OpenAI encouraged organizations interested in deploying Astra to consult the company’s technical and safety notes for full details. By combining improved desktop and browser automation, tighter alignment, enterprise-ready controls and new pricing, Astra is positioned as a model tailored to reduce friction for businesses adopting AI while addressing safety and cost concerns central to professional use.
As enterprises evaluate Astra in production settings, the model’s claimed improvements in efficiency, alignment and control will be the metrics to watch. For organizations balancing productivity gains with governance and security needs, Astra aims to make advanced automation more accessible without requiring extensive custom engineering.
Source: Read the original source
Leave a Reply