GPT-6 Astra Hits 99.9% on ARC-AGI-3, Automates End-to-End Workflows

On September 4, OpenAI unveiled its new flagship model GPT-6 Astra , describing it as the most intelligent and best-aligned model globally. The core shift is a move from merely generating text or code toward directly completing work : given a goal such as "prepare a financial analysis and turn it into a slide deck" or "build and test a game prototype," Astra autonomously decomposes the task, invokes tools, and delivers ready-to-use documents, spreadsheets, presentations, websites, or full engineering projects.

Autonomous Capabilities

Fills web forms, updates CRM records, manages calendars

Performs web search and email summarization

Analyzes scientific data, generates charts, builds websites

Installs and tests software, then troubleshoots issues appearing on screen

Benchmark Results (Near-Perfect Scores)

FrontierMath Tier 4 (high-level mathematics): 97.6%, up from GPT-5.6 Sol's 83%; reportedly assisted in solving a long-standing open problem in mathematics.

ARC-AGI-3 (abstract reasoning, dubbed the "ultimate AGI exam"): 99.9%, versus 7.8% for GPT-5.6 Sol.

ExploitBench (cybersecurity): 100% — the first model to clear OpenAI's internal "critical-level" cybersecurity capability threshold.

Terminal-Bench 4.0 (coding): 57.7%, far above GPT-5.6 Sol's 37.3%.

GPQA Diamond (science QA): 96%.

Efficiency Gains

On OSWorld 2.0, Astra scores 72.6% (GPT-5.6 Sol: 65.7%). The average time to complete a single task drops from roughly 75 minutes to 40 minutes, a 47% reduction .

Real-World Case Studies

Legora (legal tech): Used Astra for financial statement review; it examined 41 documents in minutes and detected 4 planted human errors, yielding ~40% efficiency improvement.

Playco (game studio): Applied Astra to prototype development, cutting manual bug-fixing effort by 50%.

From "answering questions" to "completing work," and from 7.8% to 99.9% on ARC-AGI-3, GPT-6 Astra marks a notable milestone in AI development. As Brockman stated: "I think it is not an unreasonable view to say we have now entered the AGI era."