SIAgentNews.com
← All news

OpenAI flags high risks as new models enable AI self-improvement

The release of GPT-6.2 Astra and Sol brings significant performance gains and cost reductions while introducing new concerns regarding autonomous AI development.

· First reported by heise.de

Ink and watercolour: picture for the story 'OpenAI flags high risks as new models enable AI self-improvement'
AI-made ink and watercolour for this story. No real people or products are shown.
36 percentreduction in average costs per benchmark task for Astra

The short version

  • GPT-6.2 Astra ranks as the second most powerful model in independent tests by Artificial Analysis.
  • GPT-6.2 Sol achieves performance levels of previous top models at roughly 15 percent of the original cost.
  • OpenAI warns that AI self-improvement could outpace human oversight and safety measures.
  • A new 'Control Room' system has been created to monitor and stop risky model workloads in research environments.
  • Astra and Sol are now available for ChatGPT Plus, Pro, Business, Enterprise, and Edu users.

Why it matters

These updates show that AI is becoming more efficient and capable of complex tasks like 3D world building and robotics. However, the shift toward AI-driven development means companies must now navigate a landscape where models can potentially improve themselves faster than humans can monitor them.

New models for efficiency

OpenAI has released two new versions of its flagship models: GPT-6.2 Astra and GPT-6.2 Sol. These models are designed to handle complex workflows more efficiently by using fewer output tokens, which helps lower the overall cost per task for businesses and developers.

Astra follows the GPT-6 Astra model released in early September, while Sol replaces the GPT-6.1 Sol version from just seven weeks ago. The company is focusing on making these tools more practical for everyday professional use, particularly in software development and scientific research.

In independent testing by Artificial Analysis, Astra reached 62 points on the Intelligence Index. This puts it in second place among the most powerful models, trailing only Claude Fable 5.2 while beating Gemini 4 Carbon.

Performance and cost gains

The new models show significant improvements in specific technical areas. Astra is noted for its ability to write higher-quality code, perform targeted testing, and create interactive 3D worlds, where it actually took the first-place spot in testing. It also shows promise in CAD simulations and robotics tasks.

For users looking for a balance between power and price, GPT-6.2 Sol is a notable update. It reaches the performance level of the previous GPT-6 Astra but at only about 15 percent of the cost per benchmark task. This allows for more complex, multi-step workflows to be handled more naturally and cheaply.

OpenAI's models currently occupy a large portion of the Pareto-Front. This refers to the set of models that cannot be improved in performance without increasing costs, or vice versa. By optimizing these metrics, OpenAI aims to provide the best possible value for high-end AI needs.

Security and self-improvement

Despite the performance gains, OpenAI has issued warnings regarding new safety risks. Both Astra and Sol have been assigned a 'Critical' risk level for cybersecurity and a 'High' level for biological and chemical capabilities. The models are now better at identifying and exploiting vulnerabilities, even as they are designed to fix them.

A major concern for the company is AI self-improvement. OpenAI has classified GPT-6.2 Astra as its first externally released model to hit the 'High' risk level for this category. This means the AI is increasingly capable of contributing to the development of its own successors.

OpenAI warns that these capabilities could allow AI development to move faster than human oversight or safety measures can keep up with. Because of this, the company is implementing stricter protections for high-risk cybersecurity requests and isolating its research environments more strictly.

New oversight systems

To manage these risks, OpenAI is introducing a central system called the Control Room. This infrastructure is designed to track model workloads within the company's research environments. It allows researchers to quickly stop any workloads that trigger specific safety concerns.

The company is also expanding its monitoring of the models' capabilities in robotics and science. While Astra's performance in coding agents has stagnated, its ability to handle complex scientific data remains a primary focus for the developers at OpenAI.

Users can access these new models through various tiers of ChatGPT, including Plus, Pro, and Enterprise. Developers can also access them via the API, where pricing for Astra has been reduced by 20 percent compared to its predecessor.

Advanced 3D and robotics focus

The new models continue to build on the strengths of their predecessors by making significant strides in the development of interactive 3D worlds. Beyond just text and code, these AI systems are being used to create complex environments and are showing promise in robotics. While the robotics benchmarks are still relatively new and focus on simple tasks, they demonstrate the models' growing versatility in physical and spatial reasoning.

Detailed API pricing and costs

For developers using the API, OpenAI has introduced specific price reductions to make high-end performance more accessible. GPT-6.2 Astra now costs 8 US dollars per million input tokens and 40 US dollars for output tokens, which represents a 20 percent decrease from the previous version. Additionally, users can benefit from a 60 percent price drop on cached inputs, which now cost only 0.40 dollars per million tokens.

Specific risk level assessments

OpenAI's internal risk assessments for these models highlight specific dangers in high-stakes domains. Both GPT-6.2 Astra and Sol have been assigned a Critical risk level regarding cybersecurity capabilities. Furthermore, they have been flagged with a High risk rating for their potential applications in biological and chemical fields, necessitating the new safety protocols and isolation measures mentioned by the company.

What to watch

  • How the Control Room system handles real-time 'self-improvement' attempts by the models.
  • Whether Astra's performance in the Coding Agent Index will improve in the next update.
  • How other companies respond to the 'High' risk rating for AI self-improvement.

Try it

Check the pricing for GPT-6.2 Astra on the OpenAI API page to see how much the input and output token costs have decreased compared to the previous version.

Sources and links

First reported by heise.de. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news