type
Post
status
Published
date
Nov 18, 2026
slug
summary
tags
category
行业动态
icon
password
OpenAI Unveils GPT-6.2 Astra and Sol: A Dual Release Reshaping AI Capabilities
OpenAI has simultaneously introduced two new models, GPT-6.2 Astra and GPT-6.2 Sol, marking a significant evolution in the company's lineup. The release positions the two models for distinct roles, with GPT-6.2 Astra targeting high-difficulty, flagship-level projects, while GPT-6.2 Sol is engineered for frequent, routine professional tasks. This dual-release strategy allows OpenAI to address both the need for peak performance in complex scenarios and the demand for cost-effective solutions in everyday operations.
Technical Performance and Specialized Gains
The new models demonstrate substantial improvements in specialized evaluations. In terms of complex agent capabilities, GPT-6.2 Astra shows a 5.2 percentage point improvement over its predecessor, GPT-6 Astra, on the Agents’ Last Exam. Notably, the cost per task has been reduced to less than one-third of the previous generation. When compared to Claude Fable 5.2, GPT-6.2 Astra leads by 5.5 percentage points on this benchmark, maintaining a single-task cost that is only 5% of the competitor's.
In the realm of scientific research, GPT-6.2 Astra has set a new standard, which OpenAI describes as a global record for scientist models. The model achieved a 9 percentage point improvement in the Terminal-Bench-Science 0.1 evaluation. This model excels in the Vals Science Index, a benchmark that requires understanding real biological data to design experiments and deduce system rules through simulation, moving beyond simple knowledge retrieval to a process closer to actual scientific work.
3D Generation and Visual Reasoning
A standout feature of this update is the advancement in 3D generation. GPT-6.2 Astra has taken the top spot on the Artificial Analysis AA-3D leaderboard with a score of 1318 Elo, surpassing Claude Fable 5.2, which recorded 1271 points. Even the lighter GPT-6.2 Sol performed well, achieving 1259 points on the same metric.
The models can now handle immersive interactive scenes and complex 3D modeling. For instance, GPT-6.2 Astra has demonstrated the ability to model a remote-controlled car, including its engine, from just photos and manuals, and to generate assembly animations. In another visual reasoning showcase, the model created an interactive 3D backyard that allows users to change perspectives and simulate sunlight at different times of the day. Furthermore, the model was observed performing mechanical design directly within CAD software, managing the intricate relationships between various components.
Cost-Efficiency and API Pricing
One of the most appealing aspects of the GPT-6.2 release is the pricing structure, which has been significantly optimized for developers. The API price for GPT-6.2 Astra has dropped by 20%, resulting in a cost of $8 per million input tokens and $40 per million output tokens. Additionally, the price for cached inputs has decreased by 60%.
GPT-6.2 Sol maintains its previous pricing, set at $2 per million input tokens and $10 per million output tokens. This positioning makes Sol an attractive option for high-volume applications where flagship-level performance is not strictly required but cost efficiency is a priority.
Professional Workflow Integration
Beyond technical benchmarks, GPT-6.2 brings notable upgrades to professional workflows, particularly in document and spreadsheet management. On the GDPval v2 evaluation, which measures real-world professional tasks, GPT-6.2 Astra improved by 6.2 percentage points over its predecessor, while Sol improved by 5.3 percentage points. Efficiency has also increased, with GPT-6.2 Astra requiring 22% fewer output tokens to complete tasks.
In practical demonstrations, GPT-6.2 Astra excels at generating complex Excel income forecasts and valuation tables that are more clear and usable than those generated by the previous version. In presentation design, the model can produce up to 8 distinct designs for a single concept, including badges, t-shirts, and embroidered backpacks, whereas previous generations were limited to a single image per team. It also produces high-fidelity market analysis slides for fashion brands, complete with categorized prices, logos, and representative product images.
Programming and Cybersecurity Capabilities
GPT-6.2 Sol is particularly notable in coding, where it has closed the gap with Astra to just 0.5 points on the FrontierCode 2.0 evaluation, while maintaining a task cost that is roughly one-fifth of the flagship model. On the LifeVal benchmark for daily tasks, Sol even outperformed the previous generation of GPT-6 Astra at a cost that was about one-seventh as high. Internal reports suggest that many OpenAI employees now use Sol as their default model for both technical and non-technical work.
Security capabilities have also seen a leap. Both Astra and Sol now reach OpenAI's highest risk tier for cybersecurity tasks. In the DeepBinary vulnerability discovery evaluation, GPT-6.2 Astra estimated total bounties exceeding $1 million for discovered vulnerabilities, a significant jump from the approximately $440,000 recorded by the previous generation.
These models are now available on ChatGPT Work, Codex, and the API for Plus and Pro subscribers.
Source: QbitAI, https://www.qbitai.com/2026/11/510707.html
- 作者:数字Baby
- 链接:myaigc.top/article/41cb7c2d-0e98-81aa-a927-d3b4d51ee6bb
- 声明:本文采用 CC BY-NC-SA 4.0 许可协议,转载请注明出处。



