OpenAI’s GPT-6 Astra: The ‘Do-It-All’ Model and the AGI Era

GPT-6 Astra: OpenAI lance son modèle « tout faire » et assume l’ère de l’AGI — AI-generated image

OpenAI rolls out GPT-6 Astra, a model that can act autonomously on a computer for complex tasks, from coding to cybersecurity. The company now embraces the AGI label while restricting access and tightening safeguards.

After the temporary suspension of Astra in August 2026 due to its cybersecurity capabilities, OpenAI relaunches its flagship model with an ‘aligned’ version and targeted restrictions.

A model that acts, not just responds

GPT-6 Astra marks a turning point: it no longer just generates text or code but executes actions directly on a computer. Via the ChatGPT desktop app (Mac, Windows, Linux), it clicks, navigates, fills forms, installs software, or updates files by analyzing screenshots. On macOS, a blue cursor signals its activity; on Windows, it takes control of the desktop. On Linux, it’s limited to shell commands and file manipulation.

To enable these features, users must allow screen recording and accessibility, then approve each program individually. Astra requests confirmation before sensitive actions and cannot control a terminal, validate admin requests, or operate ChatGPT itself. A ‘locked use’ option even lets it run on a locked Mac from a smartphone.

Record-breaking performance, but acknowledged limits

OpenAI touts unprecedented scores: 98% on FrontierMath Tier 4 (open-ended math problem-solving), 99.9% on ARC-AGI-3 (general intelligence test), and 100% on ExploitBench (exploiting software vulnerabilities). Astra also leads OSWorld, a computer control benchmark, with 72.6% success versus 65.7% for GPT-5.6 Sol, while halving the time per task.

However, these figures depend on the test ‘harness’ used. With the standard harness, shared across all models, Astra drops to 62.7% on ARC-AGI-3, while humans solve 100% of the levels. On DeepSWE (coding evaluation), it caps at 74.1%, just ahead of GPT-5.6 Sol (72.7%) and neck-and-neck with Claude Opus 5 or Gemini 3.8 Flash.

Cybersecurity: A ‘critical’ threshold and reinforced safeguards

GPT-6 Astra is OpenAI’s first model to reach the ‘critical’ threshold in cybersecurity according to its own evaluation framework. In testing, it successfully exploited unknown vulnerabilities to execute arbitrary code in hardened browsers or systems and identified two zero-day flaws. The public version now refuses advanced security tasks, and an API mechanism can halt a task if deemed risky.

Jakub Pachocki, OpenAI’s chief scientist, acknowledges that such models ‘can very well achieve a goal while acting against the intended purpose.’ CEO Sam Altman reiterates the company’s commitment to ‘slow down’ if model capabilities outpace safety and alignment progress.

Availability and cost: A gradual rollout

GPT-6 Astra is being deployed in waves: first to specialized cybersecurity customers, then to Pro, Business, and Enterprise subscribers (starting at $100/month), before reaching Plus subscribers ($20/month). It’s also available via the OpenAI API and AWS. Pro, Business, and Enterprise users get a ‘Pro’ version of Astra with extended capabilities.

On pricing, Astra costs $10 per million input tokens and $50 per million output tokens via the API, a 150% increase over GPT-5.6 Sol. Its context window reaches 1.05 million tokens, allowing it to maintain long session threads without losing initial instructions.

OpenAI embraces AGI, but benchmarks tell a different story

For the first time, OpenAI uses the term AGI (Artificial General Intelligence) to describe Astra. Company president Greg Brockman claims the model has reached this stage, defined by OpenAI as a system capable of matching human intelligence for cognitive tasks. Yet on Artificial Analysis rankings (v4.1.1), Astra still trails Claude Fable 5.1, Claude Opus 5, Muse Spark 1.3 (Meta), and Claude Fable 5 on both the Intelligence Index and Coding Agent Index.

"Anything you can do on a computer, Astra can do for you. Fast." — OpenAI, official postTranslated from French

With GPT-6 Astra, OpenAI crosses a new threshold: AI no longer just describes, it acts. The question remains whether the safeguards will be enough to contain a model whose creators acknowledge the risks. The benchmark battle, meanwhile, is far from over.