Monday: Anthropic's upcoming proactive assistant Orbit was spotted in Claude Cowork builds, showing up as a settings toggle. Google added Webhooks to the Gemini API, letting developers receive instant notifications when long-running tasks finish.

Tuesday: Gemini 3.2 Flash briefly appeared in the Gemini iOS and Android apps for some users, showing performance close to Gemini 3.1 Pro at Flash speed and cost. A Miami startup called Subquadratic launched SubQ, claiming a 12 million token context window, 52 times faster performance than the current industry standard, and costs under 5% of Claude Opus.

Wednesday: Anthropic launched Dreaming for Claude Managed Agents at their Code with Claude conference in San Francisco, letting agents review past sessions and reorganize memories to improve over time. Alongside it came Outcomes for quality-driven self-improvement, multiagent orchestration for parallel task delegation, and webhooks for job completion alerts. xAI also made its image generation model available via API, the same model that has powered over 300 million images on Grok.

Thursday & Friday: OpenAI dropped three new voice models led by GPT-Realtime-2, bringing GPT-5 level reasoning to live voice conversations for the first time. A Codex Chrome extension launched letting Codex work across browser tabs in the background. Claude launched in general availability across Excel, PowerPoint, Word, and Outlook in public beta, carrying full conversation context across all four apps simultaneously.

The Anthropic and SpaceX partnership announcement was genuinely good news for users, doubling Claude Code rate limits and removing peak hour restrictions overnight. But the fact that Anthropic needed a deal with Elon Musk's rocket company to keep up with demand says something uncomfortable about where the compute bottleneck is right now. Every major AI lab is competing for the same limited pool of infrastructure, and the winners of the AI race may end up being whoever locked in the right deals at the right time, not necessarily whoever built the best model.

The week was also unusually heavy on unverified claims and stealth appearances. Gemini 3.2 Flash showed up in the app and disappeared before Google said a word. SubQ launched with extraordinary benchmark claims and no published research to back them up. Orbit was spotted in code before any announcement. The pace at which things are leaking, appearing briefly, and vanishing is making it genuinely difficult to separate real signals from noise, and that problem is only getting worse as competition intensifies.

When every lab is racing to ship faster than they can announce, and when startups can claim to beat frontier models without publishing their methodology, the information environment around AI gets harder to trust. Not because everyone is lying, but because the pace of change has outrun the norms around verification.

The most significant thing that happened this week was not a product launch. It was Claude moving into Microsoft Office in general availability while simultaneously gaining the ability to carry context across Word, Excel, PowerPoint, and Outlook in a single conversation. Microsoft spent years and billions trying to make Copilot do exactly this. Anthropic shipped it to general availability quietly, via the AppSource marketplace, with no splashy event. The competitive dynamics in enterprise software are moving faster than the deals meant to define them.

Voice is the other story that will look more important in hindsight than it does today. GPT-Realtime-2 is the first voice model that combines real reasoning with live conversation, and Sam Altman flagged this week that voice is becoming one of the fastest growing ways people interact with AI. xAI launched a voice agent the same week. The shift from text as the primary interface to voice as a genuine option for complex tasks is happening faster than most people expected, and this week felt like the moment it became real.

Anthropic teaching agents to dream is a phrase that would have sounded like science fiction two years ago. Today it is a public beta feature you can request access to. The gap between what AI can do and what most people think it can do has never been wider, and it is widening every week.

It was an eventful week.