GPT-6 Astra: Next-Gen AGI Features & Benchmark Results

OpenAI has officially launched GPT-6 Astra, introducing an architecture designed to act as an autonomous coworker rather than a standard text predictor. By combining adaptive reasoning budgets with continuous computer control, Astra achieves state-of-the-art results across coding, scientific research, and complex professional environments.

1. Autonomous Computer Use and Operating System Navigation

Prior generations of multimodal models relied on external scripts or isolated API calls to interact with digital workspaces. GPT-6 Astra executes tasks directly inside operating systems, interpreting screens pixel by pixel, issuing precise keystrokes, and operating desktop applications like KiCad, Blender, and terminal consoles without human intervention.

On the rigorous Agents’ Last Exam benchmark, Astra recorded a 59.3% success rate, outpacing GPT-5.6 Sol (53.6%) and competitive models from Anthropic and Google. Similarly, on OSWorld 2.0, Astra scored 72.6%, demonstrating consistent precision when switching windows, executing commands, and parsing visual interface layouts.

2. Speed and Real-Time Engineering Pipelines

A key bottleneck in frontier agentic models has been latency during multi-turn self-correction. Astra implements speculative decoding and dynamic reasoning allocation, allowing it to complete multi-step software and hardware engineering pipelines in minutes rather than hours.

In an official live demonstration, Astra completed a multi-stage engineering challenge—from requirements parsing to database schema setup and verification—in 2 minutes and 54 seconds, while earlier models timed out or required manual intervention.

3. Professional Workflows, 3D CAD, and Creative Assets

Astra demonstrates significant practical capability in professional knowledge work. In 3D computer-aided design, the model achieves 95.9% on the BenchCAD evaluation, up from 83.3% on GPT-5.6 Sol. It can reconstruct complex mechanical objects from multi-view renders by writing functional parametric Python code.

When building interactive web applications and digital assets, Astra avoids generic template outputs. It renders fully styled components, handles state management, and integrates responsive layouts seamlessly.

4. Coding, Science, and Abstract Reasoning

On developer evaluations, Astra set new records on Terminal-Bench 4.0 with a 57.9% score, substantially higher than GPT-5.6 Sol (37.3%) and Claude Fable 5.1 (55.8%). In mathematical and scientific reasoning, Astra scored 97.6% on FrontierMath Tier 4 and 96.0% on GPQA Diamond.

According to the ARC Prize Foundation, Astra reached human parity on 96% of ARC-AGI-3 levels, marking a significant milestone in general abstraction and pattern transfer.

5. Cybersecurity, Safety, and Hallucination Control

Autonomous agentic systems present novel safety challenges, including unauthorized network access and tool misuse. OpenAI evaluated Astra against specialized honeypots and red-teaming benchmarks. On the ExploitGym honeypot, Astra achieved a 0.0% unauthorized exploit rate, compared to 48.2% on GPT-5.6 Sol.

In alignment evaluations measuring misaligned outcomes during desktop execution, Astra achieved critical safety improvements across key threat models:

6. Official Benchmark Evaluation Tables

The tables below present full test scores across standard academic, coding, professional, security, and long-context evaluations as published by OpenAI and verified by independent testing labs including Artificial Analysis and Epoch AI. Comparison models include Claude Opus 5 by Anthropic and Gemini 3.8 Flash by Google DeepMind.