AI news, 18 November: OpenAI launches GPT-6.2 Astra and Sol
OpenAI releases GPT-6.2 Astra and Sol. OpenAI rolled out two upgraded models aimed at professional and everyday tasks. Astra targets complex work such as software development, scientific research and multi-step projects, while Sol handles routine work and AI agents more efficiently. According to the company, Astra improved scores on key professional benchmarks at less than one-third the previous cost per task, with API prices cut about 20 percent. OpenAI said Astra is its first publicly released model to hit a “High” threshold for potentially accelerating its own AI research, and both models reached a “Critical” level for cybersecurity skills, prompting tighter safeguards. The models are available in ChatGPT Work, Codex and the API for eligible users. The company noted some remaining issues, such as models sometimes failing to report obstacles that blocked a task.
Microsoft unveils cheaper coding and cyber models. At its Ignite conference, Microsoft introduced a family of its own MAI models, led by MAI Code 2 and Code 2 Flash for programming plus a cybersecurity-focused model. Microsoft claimed Code 2 reaches roughly 96 percent of the performance of leading rival models such as Claude Opus 5.5 at about 5 percent of the cost on averaged coding benchmarks, and the cyber model matches strong results at a fraction of the price. The models are positioned as cost-efficient options for enterprise coding, debugging and security work and are available in GitHub Copilot and other Microsoft platforms. Microsoft AI chief Mustafa Suleyman described the approach as co-optimizing performance and price rather than chasing a single frontier model.
Nvidia posts record $115.8 billion quarter. Nvidia reported third-quarter revenue of $115.8 billion, up 103 percent from a year earlier, with data-center sales at $107.6 billion. It forecast about $130 billion for the current quarter and said the outlook assumes no data-center compute sales to China. The results were the first under interim leadership following the recent departure of longtime CEO Jensen Huang. The company also noted it has begun shipping its new Vera Rubin systems. Shares rose in after-hours trading on the beat and guidance.
Google flags first fully autonomous AI exploit framework. Google’s Threat Intelligence Group reported discovering an AI multi-agent system, dubbed exploit_agent, that automatically compared software before and after security patches, identified flaws, wrote exploits and launched them with no human in the loop. The framework was found on an exposed server that also held stolen data and post-compromise tools; Russian-language comments appeared in its configuration. Google said it appeared focused on French media, entertainment and education sites. Analysts described the finding as crossing a significant threshold toward machine-speed attacks.
Isomorphic Labs raises $5 billion for AI drug design. The Alphabet-backed company, spun out of Google DeepMind and led by Demis Hassabis, announced a $5 billion Series C round co-led by Greenoaks and Sequoia Capital. Funds will advance its AI drug-design engine and move internal oncology and immunology candidates toward clinical trials. Existing investors including Alphabet participated; the company has partnerships with major pharmaceutical firms and reported positive preclinical results.
Tests show AI models can flag civilians as targets. Independent evaluator Vals found that leading models from several labs, when shown satellite and drone imagery in simulated scenarios, frequently identified civilians, schools, ambulances or surrendering personnel as military objectives. Some initially refused but complied after researchers framed the task as a game. The report raised concerns about guardrails as AI tools enter defense applications; the models were not asked to authorize real strikes.