Artificial Intelligence

OpenAI unveils GPT-6.2 Astra and Sol models

OpenAI's new GPT-6.2 Astra model beat its predecessor on document drafting, data analysis and scientific tasks, while the everyday GPT-6.2 Sol delivers more work without a price increase.

OpenAI unveils GPT-6.2 Astra and Sol models

As the AI race continues, OpenAI has unveiled its much-anticipated new models, GPT-6.2 Astra and Sol.

According to OpenAI's announcement, GPT-6.2 Astra was developed for professional work and GPT-6.2 Sol for everyday use. Response quality has improved in both models, while usage costs have decreased rather than increased.

Advertisement

GPT-6.2 Astra was developed for professional tasks

According to the company, GPT-6.2 Astra outperformed its predecessor GPT-6 Astra by 5.2 percent on the Agents' Last Exam, a test that measures multi-step professional tasks. This result was achieved while completing the tasks with less than a third of the spending. On the same test, when compared with Claude Fable 5.2, Astra came out 5.5 percent ahead, at a cost of roughly 5 percent of the rival's.

For scientific work, OpenAI stated that on the Terminal-Bench-Science 0.1 test, Astra delivered 9 percent better results than the previous version, and its usage cost was also one-third lower. The model was also reported to achieve the best result on the Vals scientific index. The company further says there is a quality improvement in presentations, financial statements, legal documents and analysis reports. On the internal GDPval v2 test, Astra was reported to show a 6.2 percent improvement over its predecessor, while output token usage dropped by 22 percent.

Image source:

It surpassed all its rivals

In the field of 3D modeling, OpenAI shared an example in which Astra simulates the movement of a car chassis on a mountain road. It was described as creating a 3D model, writing a calculation block in Python, and animating the result in Blender. On the AA-3D test, Astra scored 1318 Elo points; Claude Fable 5.2 scored 1271, and GPT-6.2 Sol scored 1259.

Advertisement

The everyday model GPT-6.2 Sol was introduced as smarter and more efficient than GPT-6.1 Sol. The model supports multi-step agent workflows. On Cognition's FrontierCode 2.0 test, Sol trailed Astra by only 0.5 percent, at about one-fifth of the cost. OpenAI noted that one area where Sol performs roughly seven times better than its previous version is tasks users assign to AI in their free time.

OpenAI unveils GPT-6.2 Astra and Sol models

The company reported that both models fall into the "critical" risk category for cybersecurity and the "high" risk category for biology and chemistry. Astra became the first model OpenAI has made publicly available to reach the "high" level of self-improvement capability. For this reason, restrictions on potentially dangerous requests have been tightened and the models' actions are being monitored more closely.

OpenAI unveils GPT-6.2 Astra and Sol models

In terms of access, GPT-6.2 Astra and Sol are offered to ChatGPT Plus, Pro, Enterprise and Edu users through ChatGPT Work and Codex, and developers can access them via the API. Astra is priced at $8 per million input tokens, $0.40 for cached input and $40 for output. Sol's prices remain the same as the previous version: $2, $0.10 and $10 respectively.

Advertisement

OpenAI unveils GPT-6.2 Astra and Sol models

Users in Turkey have recently turned to Claude. We'll see whether ChatGPT's user count increases after these models, so let's wait and see.

Advertisement

This article was translated from Turkish. Read the original article in Turkish.

Follow us on Google News