SIAgentNews.com
← All news

OpenAI launches GPT-6.2 Astra and Sol with lower API pricing

The new models offer a performance-to-cost advantage for professional science and everyday tasks while hitting a new threshold for AI self-improvement.

· First reported by The Decoder

Claymation: picture for the story 'OpenAI launches GPT-6.2 Astra and Sol with lower API pricing'
AI-made claymation for this story. No real people or products are shown.
20 percentlower API pricing for GPT-6.2 Astra

The short version

  • GPT-6.2 Astra is designed for high-demand fields like science, law, and finance with a 20% lower API price than its predecessor.
  • GPT-6.2 Sol targets everyday tasks and offers a significant cost advantage, costing roughly one-fifth of Astra for similar coding performance.
  • Astra leads the Vals Science Index and sets a new high on the Terminal-Bench-Science benchmark.
  • Astra is the first externally released model to hit OpenAI's 'High' threshold for AI self-improvement.
  • The models are now available to Plus, Pro, Enterprise, and Edu users across ChatGPT Work and Codex.

Why it matters

These releases signal a shift toward 'cost-performance' optimization, making high-end reasoning more accessible for businesses. For users, it means more capable tools for complex workflows like 3D modeling and scientific research at a lower price point.

Two models for different needs

OpenAI has introduced two new language models, GPT-6.2 Astra and GPT-6.2 Sol, to cater to different user requirements. Astra is built for demanding professional work in specialized fields such as law, finance, and accounting. It is designed to handle professional conventions more reliably, such as pulling layouts and typography from specific templates for presentations.

In contrast, GPT-6.2 Sol is geared toward quick, everyday tasks and general use. While it is less specialized than Astra, it provides a versatile experience for common queries. Both models are now available to Plus, Pro, Enterprise, and Edu users within the ChatGPT Work and Codex platforms.

The company is emphasizing a better price-to-performance ratio with these releases. Astra's API pricing is 20 percent cheaper than its predecessor, while Sol maintains the pricing of the previous version. This strategy aims to provide high-quality reasoning while reducing the operational costs for developers and businesses.

Benchmark gains and science focus

The new models show notable improvements in scientific and 3D design benchmarks. Astra leads the Vals Science Index, which measures how well models interpret biological data and run experiments. It also scored 9 percentage points above its predecessor on the Terminal-Bench-Science benchmark, which tests scientific workflows in a terminal environment.

In the realm of 3D design, Astra took first place on the AA-3D benchmark, where human judges compare interactive experiences. It also demonstrated improved computer use capabilities, such as building a 3D car model and writing a physics solver in Python. These capabilities suggest a stronger grasp of complex, multi-step instructions involving external tools.

However, the models faced mixed results in other areas like coding and legal research. While Astra doubled its predecessor's pass rate on some legal benchmarks, it ranked lower overall in legal research tests. Additionally, both models saw flatlined or even slightly declining scores on certain coding agent indexes compared to previous versions.

The cost-performance frontier

According to the Artificial Analysis Intelligence Index, GPT-6.2 Astra ranks second overall, trailing only Anthropic’s Claude Fable 5.2. Despite this, OpenAI's model dominates the Pareto frontier, which is the line where no other model is both cheaper and better. This means Astra provides a highly competitive balance of intelligence and affordability for high-end tasks.

Sol offers an even more extreme example of this efficiency. It nearly matches the original Astra's performance on some tasks while costing only about 15 percent of the price per task. This makes Sol one of the cheapest models in the top 10 rankings provided by Vals AI, appealing to users who need speed and versatility without the premium price tag.

OpenAI is also highlighting token efficiency as a key metric for these updates. Astra uses the fewest output tokens in the top 10 rankings, averaging 91,000 per task. By using fewer tokens to achieve similar results, the company helps users reduce the costs associated with long-form generation and complex reasoning steps.

Self-improvement and safety risks

A significant technical milestone for GPT-6.2 Astra is reaching the 'High' threshold for AI self-improvement. This means the model can meaningfully speed up AI research, which OpenAI compares to giving every researcher an extra, highly skilled engineer. The company explicitly identifies this as a risk because models that can build future AI systems could outpace human oversight.

To manage these risks, OpenAI is implementing stricter controls in research environments and using a monitoring system called Control Room. This system is designed to track research tasks and shut them down if unauthorized behavior is detected. They are also expanding monitoring to examine how models reason and act in real-time to flag potential issues.

The company also noted that the new models hit the 'Critical' threshold for cybersecurity in their Preparedness Framework. They are adding stricter safeguards for risky cybersecurity requests to prevent abuse. While these measures may limit legitimate security testing by researchers, they are intended to make it harder for malicious actors to exploit the models for finding vulnerabilities.

What to watch

  • How the 'High' threshold for AI self-improvement will impact the speed of future model releases.
  • Whether the new cybersecurity safeguards will hinder legitimate security researchers using OpenAI Daybreak.

Try it

Check the Artificial Analysis Intelligence Index to see how Astra ranks against other models in cost-performance.

Sources and links

First reported by The Decoder. 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