OpenAI said it will restrict the most advanced cybersecurity capabilities of its new artificial intelligence model, Astra, aiming to prevent a repeat of the Hugging Face incident in which an AI model carried out hacking on its own. Rival Anthropic, meanwhile, plans to lower the token prices charged for using its major models, including the newly released Fable 5.1. As competition among AI models intensifies, both companies are moving to minimize vulnerabilities.
OpenAI said in a blog post on the 1st that it will soon unveil Astra. The company did not specify a release date but stressed that the ability to perform advanced cybersecurity tasks will initially be available only to a limited group.
OpenAI halted part of its work on Astra last month. The step was taken to build in stricter safeguards after the model was found to demonstrate considerable capability in cybersecurity tasks. The company said it considers Astra to have reached a "critical cybersecurity threshold," meaning the model can identify and exploit zero-day vulnerabilities on its own. A zero-day vulnerability is a software flaw that attackers exploit before a patch is available.
The company also strengthened measures to prevent misuse of Astra in cybersecurity-related activity. These include functions that monitor and automatically block unauthorized behavior during internal deployment. Access will later be expanded to more users through a program called Daybreak Blue, which limits use to the intended purpose of cybersecurity.
Anthropic cut its token prices the same day. The company unveiled Fable 5.1, a new model with improved performance on coding and scientific tasks and lower operating costs. Users can see cost savings of about 25% on typical tasks and up to 45% on tasks that reference large amounts of existing context.
Anthropic said the adjustment "will significantly reduce the total cost of running the model." The company said Fable 5.1 performs better on long and complex programming tasks and has improved at supporting scientific work such as designing experiments, simulating experimental results and reading information from complex charts and tables.
Claude Opus 5, released last month, was priced at half the level of Fable 5, at $25 per million output tokens. Anthropic also withdrew a planned price increase for Sonnet 5 that had been scheduled for this month. Sonnet 5 launched at lower prices than its predecessor Sonnet 4.6, at $2 per million input tokens and $10 per million output tokens.
Bloomberg said companies are willing to give up revenue to lower AI costs because low-cost open-weight models from China are rapidly improving, and that the trend is putting pressure on U.S. model developers such as Anthropic.
