
The Zhitong Finance App learned that after putting on hold the previous model update for nearly two months to establish stricter security measures for this technology, OpenAI is launching a new version of its most advanced artificial intelligence (AI) model, Astra. OpenAI said on Tuesday that GPT-6.2 Astra aims to be the company's most capable model at present, capable of complex professional projects, including building financial models and carrying out scientific research work.
OpenAI said that judging from specific applications, the GPT-6.2 Astra upgrade focuses not only on improving the accuracy of the model's question answering questions, but also on further improving its ability to handle complete professional workflows. OpenAI said that the new model's ability to generate documents in the fields of finance, law, consulting, sales and accounting has improved, can better understand the templates, data, and industry specifications provided by users, and generate presentations, spreadsheets, and research reports that are more in line with the actual work requirements of professionals.
In the internal GDPval v2 benchmark, the model achieved leading results in multi-step real project tasks involving consulting, finance, and design, increasing the score by 6.2 percentage points compared to the previous generation GPT-6 Astra, while reducing the number of tokens exported by 22%. In the legal professional task test Harvey LAB-AA, the GPT-6.2 Astra's pass rate was more than double that of Anthropic's Fable 5.2 model, and the cost was only about one-sixth of the latter.
In terms of financial analysis, the model also showed strong competitiveness. In Artificial Analysis' AA-Analystent testing, models need to extract data from real financial reports and spreadsheets, perform calculations, and answer professional questions. GPT-6.2 Astra achieved an accuracy rate of 62.5% in each of the five independent tests in a test involving California Medicaid spending, a record high for this test.
In addition, OpenAI also emphasized that the new model can more accurately follow the presentation template, including details such as page layout, color scheme, font, and font size, and improve the generation quality of financial forecasting models, valuation workbooks, and investment memorandums. These capabilities are expected to make them more widely used in fields such as investment banking, asset management, consulting, and corporate finance.
Anthropic and OpenAI are in a fierce competition, and both are striving to develop more advanced AI models to attract a wider corporate customer base. OpenAI's decision to launch a new version of Astra comes just a few days before Anthropic's much-anticipated Wall Street listing launch. After postponing the original public listing plan until next year, OpenAI is currently seeking a new round of financing at a valuation of 1.25 trillion US dollars.
The release of this model is expected to further intensify external scrutiny of security risks in cutting-edge AI systems. Earlier, both leading AI startups had stated that after reporting a series of security incidents, it was reasonable to appropriately slow down the progress of technology development. In these events, technical test versions of the two companies visited third-party platforms on the internet.
In September, OpenAI announced the suspension of the release of GPT-6.1 Astra because the model failed to meet the security standards set by the company in terms of complying with the scope of user authorization, avoiding unauthorized operations, and transparently explaining work progress to users. Earlier, OpenAI also revealed a security incident where the model agent broke through the original quarantine environment and accessed the external Internet during internal testing. Related incidents prompted the company to suspend some cutting-edge reinforcement learning training and further strengthen network isolation, container and virtual machine access control, and operation monitoring mechanisms in the research environment.
OpenAI said that cutting-edge reinforcement learning training has now been resumed, but the training process uses cached network content instead of real-time internet access to reduce the risk of the model connecting to an external network without authorization.
At the same time, the cybersecurity capabilities of next-generation models also complicate regulatory and security issues. OpenAI previously stated that the Astra series models already have the ability to discover previously unknown software vulnerabilities and develop methods to exploit corresponding attacks with the support of appropriate tools and permissions. Therefore, the company needs to strengthen security testing before model release and continuously monitor model behavior during actual deployment.
This also means that as AI agents gradually shift from answering questions to independently perform multi-step tasks, how to ensure that models always comply with the scope of user authorization and avoid unauthorized actions will become an important issue that must be solved in the commercialization process of AI enterprises.