OpenAI launches GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper models

Key Takeaways

OpenAI launches three new audio models for real-time translation

OpenAI announced three new audio models available through its API that improve how AI handles sound and language. The first model offers real-time translation across 70 languages, allowing users to speak in one language and hear the translation instantly. The second model creates intelligent voice agents that can reason and take actions, such as pulling up meeting details in a CRM system. The third model ensures seamless integration by capturing audio directly from microphones without needing extra editing. These tools aim to break down language barriers and make voice interaction more natural for global communication.

OpenAI releases GPT-Realtime models for smarter voice interactions

OpenAI introduced three new real-time voice models to its API: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. GPT-Realtime-2 features GPT-5-class reasoning capabilities and a larger context window of 128K tokens to handle complex tasks. GPT-Realtime-Translate supports live speech translation from over 70 input languages into 13 output languages. GPT-Realtime-Whisper provides ultra-low latency transcription for instant live captions. Companies like Zillow and Deutsche Telekom are already using these models to improve call success rates and cross-language customer support.

New OpenAI models enable voice agents to reason and act

OpenAI launched a new generation of real-time voice models that allow AI to listen, reason, translate, and take action as conversations happen. The release includes GPT-Realtime-2 for complex reasoning, GPT-Realtime-Translate for multilingual support, and GPT-Realtime-Whisper for live transcription. These models support three emerging patterns: voice-to-action for completing tasks, systems-to-voice for providing spoken guidance, and voice-to-voice for cross-language communication. Developers can test these tools in the OpenAI Playground to build more natural voice experiences.

AI cannot yet generate reliable hardware from simple specifications

An expert argues that AI cannot currently generate reliable hardware designs directly from specifications without human expertise. Past attempts to automate hardware design failed due to issues with intellectual property models and a lack of accurate performance models. While AI might help generate specific models for analysis, it cannot yet ensure the resulting hardware is secure, energy-efficient, or cost-effective. The article notes that modern systems are software-centric, making the idea of software engineers creating hardware without knowledge impractical at this time.

Global AI adoption rises as usage reaches 17.8 percent in 2026

Microsoft reported that global AI usage increased to 17.8 percent of the working-age population in the first quarter of 2026. The United Arab Emirates leads the world with a 70.1 percent adoption rate, while the United States moved up to 21st place with 31.3 percent usage. AI coding tools have driven a 78 percent increase in software code production and a rise in software developer employment. However, a gap remains between the Global North at 27.5 percent usage and the Global South at 15.4 percent.

Duke students build real AI tools to help local newsrooms

Students at Duke University built five practical AI tools to help local media organizations automate tedious tasks. The class emphasized learning from failure and focusing on specific problems rather than just the latest technology. Projects included a grocery price tracker, a meeting analyzer for city council agendas, and a tool to summarize daily newsletters. The students found that traditional coding techniques were often better than generative AI for tasks like sorting and summarizing, reducing the risk of errors.

NASA deploys first AI geospatial model in orbit on satellites

NASA and IBM successfully deployed the Prithvi geospatial AI foundation model aboard the International Space Station and a satellite in orbit. This is the first time such a model has been used in space to analyze Earth data. The model, trained on 13 years of satellite data, can detect floods, clouds, and fire burn scars before the images reach the ground. Researchers used the open-source model to test its performance on two different platforms, proving that AI can perform advanced analysis in orbit.

Mozilla uses AI to hunt security bugs in Firefox browser

Mozilla ran an AI-powered system to find security bugs in the Firefox browser codebase over several months. The AI agent, powered by Claude Opus 4.6, discovered several critical flaws including a 15-year-old HTML bug and a race condition that could allow hackers to manipulate data. The system operated in isolated virtual machines to ensure safety, and any code changes made by the AI were only used to generate bug reports. Mozilla plans to integrate this scanning into their continuous development process.

Security becomes critical as AI factories scale up operations

The rapid growth of AI factories is forcing companies to redesign their security infrastructure from the ground up. These new facilities require complex power and cooling systems while storing sensitive data like model weights and secrets. Experts warn of new threats where attackers might turn trusted AI agents against an organization. Companies must now ensure that AI agents operate with the minimum necessary access and that their actions are fully monitored and observable.

Penn State hires first vice provost for AI in Pennsylvania

Penn State has hired Vasant Honavar as its first vice provost for artificial intelligence, a role likely unique in Pennsylvania. Honavar will oversee the university's AI strategy and promote ethical AI innovation across teaching and research. He brings over three decades of experience in AI and previously helped develop Penn State's AI degree curriculum. The university also appointed Crystal Ramsay as assistant vice provost for AI to lead the new AI Center of Excellence for Teaching and Learning.

Colorado bill aims to ensure transparency in AI decision making

Colorado lawmakers are considering a bill to ensure AI systems are transparent and accountable, especially for people with disabilities. The proposed law requires companies to notify consumers when AI makes consequential decisions about jobs, housing, or healthcare. It also mandates that humans review automated decisions to prevent discrimination based on historical biases. The bill seeks to protect privacy and ensure that AI tools do not exclude vulnerable groups from essential services.

Organizations must adopt crypto agility to prepare for quantum threats

IBM engineer Jeff Crume explains that organizations need 'crypto agility' to prepare for the future threat of quantum computers breaking current encryption. Cryptographic algorithms have a limited lifespan and will eventually become obsolete as technology advances. Crypto agility allows companies to quickly update their security protocols when new vulnerabilities are found or new standards emerge. A four-step process of discovery, evaluation, prioritization, and remediation is recommended to manage this transition effectively.

Sources

• OpenAI Unveils Three New Audio Models in API
• OpenAI's New Voice API Models
• Advancing voice intelligence with new models in the API
• Hardware From Specifications Using AI
• The state of global AI diffusion in 2026 - Microsoft On the Issues
• In this Duke journalism class, failure was part of the assignment. It led to real AI tools for local news.
• NASA’s Prithvi Becomes First AI Geospatial Foundation Model In Orbit
• What Mozilla learned running an AI security bug hunting pipeline on Firefox
• AI infrastructure security becomes key as AI factories scale
• Penn State names a vice provost for AI, likely a first for Pennsylvania amid the state's tech surge
• Why Colorado needs accountability and transparency in AI decision making
• The Rise of Crypto Agility: Preparing for the Quantum Era