Voice AI has been a weak spot across the industry for a while. Models that can talk have generally not been models that can think. GPT-Realtime-2 is OpenAI's attempt to fix that gap entirely. It brings the same reasoning capability as GPT-5 into live spoken conversations, meaning you can have a real back and forth with an AI that is actually working through your problem in real time, not just pattern-matching a response. It handles interruptions naturally, uses tools mid-conversation, recovers when things go sideways, and remembers the full context of your conversation as you go.

On audio benchmarks it scores 96.6% accuracy, which puts it at the top of every leaderboard it has been tested on. Two companion models launched alongside it: GPT-Realtime-Translate for live translation across 70 languages, and GPT-Realtime-Whisper for real-time captions and transcription as words are spoken. All three are in the API today.

Sam Altman noted that voice is becoming one of the most popular ways people interact with AI, especially when they have a lot of context to share quickly and typing feels like too much friction.

On the Codex side, OpenAI launched a Chrome extension that lets Codex work across multiple browser tabs in the background without hijacking your browser. It can test web apps, gather context across different tabs, and update things like CRMs and Google Docs simultaneously. The key detail is that it works in parallel without taking over your screen, which has been the frustrating tradeoff with browser automation tools until now. Available on macOS and Windows, though EU and UK users will have to wait a little longer.

Anthropic launched general availability for Claude in Excel, PowerPoint, and Word today, with Outlook joining in public beta. The full suite is available now via Microsoft AppSource for paid plan users, and the reaction from early users has been enthusiastic enough that the jokes are already writing themselves. One comment that captured the mood: "your coworker who is really good at Excel is finished."

What makes this feel like more than a typical AI integration is how the context works across apps. You start drafting a proposal in Word, switch to Excel to pull in the numbers, then build a PowerPoint from the same conversation, and Claude remembers everything across all three without you having to re-explain anything. In Word it suggests tracked changes rather than rewriting directly, so you stay in control of the document. In Excel it highlights the relevant cells without breaking your formulas. In PowerPoint it builds slides that match your existing templates rather than starting from a generic layout.

Microsoft Copilot has been trying to do something like this for years. Claude just launched it in general availability. The gap between what Copilot promised and what it delivered has been one of the most talked-about disappointments in enterprise software, and Anthropic is walking straight into that conversation.

Grok Voice Think Fast 1.0 launched today as a voice agent built specifically for business workflows that need to actually work under pressure. The pitch is that most voice AI falls apart when conversations get complicated, when there are multiple steps involved, when background noise is an issue, or when the model needs to call several tools in sequence. Think Fast 1.0 is designed to handle all of that without dropping the thread. It is available to try for free on xAI's developer console right now, which is a low-friction way to see if it lives up to the name.

On top of that, Grok Build, xAI's upcoming coding desktop app, has been spotted ahead of its official release. Screenshots are already circulating and it looks close to ready. It is coming to macOS, Windows, and Linux and will support planning mode, plugins, and a built-in browser, with the ability to spin up development servers and work directly with your code history. No release date confirmed yet, but the build activity suggests it is not far off.