Economy/Home · Economy

OpenAI Unveils GPT-6 Astra: What Has Changed from GPT-5.6 Sol?

With OpenAI's unveiling of its next-generation artificial intelligence (AI) model, "GPT-6 Astra," attention is focused on the performance gap with the existing GPT-5.6 Sol. According to OpenAI, significant changes have emerged not only in simple answer accuracy but also in real-world task capabiliti

Wooil Shim
Staff Reporter
12 min read
OpenAI Unveils GPT-6 Astra: What Has Changed from GPT-5.6 Sol?
CBC News

With OpenAI's unveiling of its next-generation artificial intelligence (AI) model, "GPT-6 Astra," attention is focused on the performance gap with the existing GPT-5.6 Sol. According to OpenAI, significant changes have emerged not only in simple answer accuracy but also in real-world task capabilities, including computer use, coding, scientific research, sustained long-horizon tasks, and cybersecurity.

■ Directly Operating Computers: "Fast, Accurate, and Safe"

The most notable difference is the ability to directly operate computers. GPT-6 Astra is designed to perform multi-step tasks such as filling out online forms, managing schedules, updating CRM records, conducting web research, writing documents, analyzing data, and building websites. OpenAI stated that Astra has improved over previous models in the speed, accuracy, and safety of computer use.

The gap also appeared in actual evaluations. On OSWorld 2.0, GPT-6 Astra scored 72.6%, outperforming GPT-5.6 Sol's 65.7%. In a task-time simulation released by OpenAI, Astra achieved its performance in about 40 minutes, while GPT-5.6 Sol took about 75 minutes. This means the model has evolved beyond simply solving problems better—it now completes the same work faster in real computing environments.

■ Coding and Software Engineering at "Highest Level Ever"…Long-Task Context Retention Also Improved

The difference was substantial in coding as well. On Terminal-Bench 4.0, GPT-6 Astra scored 57.9%, while GPT-5.6 Sol managed only 37.3%. On DeepSWE v1.1, Astra recorded 74.1% versus Sol's 72.7%. In an internal evaluation on database migration, the gap was wider: Astra scored 63.9% compared to Sol's 42.7%. OpenAI described GPT-6 Astra as its strongest software engineering model to date.

Notably, the way it maintains context in long coding tasks has also changed. Previous models, including GPT-5.6 Sol, mainly compressed and summarized past content as tasks dragged on, which risked losing details such as the reasons for earlier failed fixes or specific test results. GPT-6 Astra, by contrast, is designed to maintain notes across multiple context windows in Codex and to retrieve previous work records. This strengthens its ability to retain earlier context during long debugging sessions or large-scale code modifications.

■ Abstract Problem-Solving and Scientific Reasoning Show the Widest Gap

The gap is even more pronounced in abstract problem-solving. On ARC-AGI-3, GPT-6 Astra scored 99.9%, while GPT-5.6 Sol scored 7.8%. On ARC-AGI-2, Astra recorded 95.0% versus Sol's 92.5%.

Astra also stepped up in science. On GPQA Diamond, which evaluates graduate-level scientific reasoning, Astra scored 96.0% compared to GPT-5.6 Sol's 94.6%. On FrontierMath Tier 4, Astra recorded 97.6% versus Sol's 83.0%.

■ Improved Business Document Creation and 3D Design Capabilities

Business document creation capabilities have also changed. OpenAI explained that GPT-6 Astra has improved abilities to create documents, presentations, and spreadsheets while preserving existing templates. It was reportedly trained to extract only the context relevant to the task and reflect it in outputs, rather than repeatedly including unnecessary information. On BenchCAD, which evaluates 3D design capability, Astra scored 95.9%, far exceeding GPT-5.6 Sol's 83.3%.

■ Cybersecurity: Discovered Two Zero-Day Vulnerabilities…Some Restrictions in Service

Cybersecurity performance is one of the most sensitive areas with major changes. On ExploitBench, GPT-6 Astra scored 100%, while GPT-5.6 Sol scored 78.5%. On ExploitGym, Astra recorded 42.4% versus Sol's 30.3%. On SRE-Bench, the gap was large: Astra scored 88.0% versus Sol's 55.9%. OpenAI said that during evaluations, Astra discovered and exploited two previously unknown zero-day vulnerabilities. At the same time, considering the potential for misuse, it has placed restrictions on certain cybersecurity tasks in actual services, such as creating advanced attack proof-of-concept code.

■ Safety and Instruction Following: Error Rate Improved from 22.0% to 2.4%

Safety and user instruction following also differ from GPT-5.6 Sol. In OpenAI's internal computer-use safety evaluation, where lower error rates are better, GPT-6 Astra recorded 2.4%, while GPT-5.6 Sol recorded 22.0%. With automatic review features applied, Astra dropped to 1.8% and Sol to 4.3%. In the ExploitGym honeypot evaluation, GPT-5.6 Sol acted outside permitted target boundaries 48.2% of the time, while Astra recorded 0%. OpenAI presented this as evidence of improved ability to maintain user intent and task scope.

■ From "AI That Answers Well" to "Agentic AI That Gets Work Done"

In summary, the shift from GPT-5.6 Sol to GPT-6 Astra is less about simple improvements in answer quality and more about strengthening the ability to carry out real tasks to completion. Whereas GPT-5.6 Sol was a high-performance AI model that answered complex questions and wrote code, GPT-6 Astra moves closer to an agentic AI that directly uses computers, browsers, and development tools to carry out multi-step work and deliver finished results.

OpenAI plans to roll out GPT-6 Astra to select organizations first, then expand access to ChatGPT Plus, Pro, Business, and Enterprise users. It will also be available via the OpenAI API, Microsoft Azure, and AWS Bedrock, with the API model name "gpt-6-astra."

Ultimately, the biggest difference between GPT-5.6 Sol and GPT-6 Astra can be summarized as a shift in the standard of competition—from "how well it answers" to "how complex a task it can actually take on and complete."

[This article was written with AI assistance based on official GPT-6 Astra materials released by OpenAI. Performance figures and comparison results in the article are based on benchmarks and internal evaluations presented by OpenAI, and perceived performance may vary depending on actual service environments and usage conditions. This content is not intended to recommend the use of any specific AI service or guarantee its performance. AI assisted in part.]

Wooil Shim
Staff Reporter

CBC Globe publishes verified stories with editorial review, source checks, and tenant-specific publication standards.