Monday: OpenAI quietly pushed Spud as a stealth update inside GPT-5.4 Pro. Anthropic shipped Live Artifacts in Cowork, letting Claude build dashboards that refresh automatically with live data. OpenAI also launched Chronicle in Codex, a feature that takes periodic screenshots of your Mac to build a memory of what you have been working on.

Tuesday: ChatGPT Images 2.0 launched and immediately topped every category on the Image Arena leaderboard and became the first image model with thinking capabilities.

Wednesday: OpenAI launched workspace agents in ChatGPT, purpose-built agents that run on a schedule and keep working in the background without supervision. Google launched their own enterprise agent platform the same afternoon.

Thursday: GPT-5.5 officially dropped with meaningful gains in coding, math, and web browsing at the same speed as 5.4. Anthropic launched persistent memory for Claude Managed Agents in public beta. OpenAI shipped Auto-review in Codex, letting agents work through long tasks without asking permission at every step.

Friday: DeepSeek dropped V4-Pro and V4-Flash with no announcement, fully open source, scoring within striking distance of Opus 4.7 at a fraction of the cost. Anthropic refreshed Claude Code on the web with a new UI and cross-device continuity.

The Anthropic pricing story started as a X callout and turned into something more revealing. Users noticed Claude Code had been quietly removed from the Pro plan for some new signups, with no email, no blog post, just a pricing page edit. Upgrading to keep access meant jumping from $20 to $100 a month. Anthropic's head of growth stepped in to say it was a small test affecting 2% of new users. OpenAI's Thibault Sottiaux showed up in the replies shortly after to confirm that Codex remains on free and $20 plans with no asterisk. The timing could not have been worse.

But the story underneath the story is worth paying attention to. Anthropic is running some of the most capable AI in the world, and the economics of doing that at scale are getting complicated. The way people use a Claude subscription has changed fundamentally since Max launched, with agents running for hours, Cowork automating entire workflows, and Claude Code handling tasks that would have required a team six months ago. The current pricing was not built for this. Anthropic said as much. The question is how they thread the needle between covering costs and keeping the trust of a community that chose them partly because they seemed different from the companies that nickel and dime their users.

Two things defined this week more than anything else. The first is $65 billion. Amazon committed $25 billion to Anthropic earlier this month. Google committed $40 billion on Friday. Both companies run competing AI models but neither is making a bet that Anthropic wins the race. What they are buying is compute dependency, the guarantee that when Anthropic's models run, they run on their infrastructure, and every query, every API call, every product built on top of Claude generates revenue for them. It is infrastructure strategy dressed as belief in a mission.

The second thing is DeepSeek. GPT-5.5 landed this week as the culmination of two years of research, and it is genuinely impressive. DeepSeek dropped V4 the same week, open source, for a fraction of the price, within striking distance on most benchmarks that matter to developers. The gap between American and Chinese AI that felt decisive twelve months ago now looks like a rounding error. And unlike GPT-5.5, which sits behind a subscription, anyone in the world can download DeepSeek V4 right now and run it on their own hardware.

The race for AI dominance is being run on two tracks simultaneously. One track is about who builds the most capable model. The other is about who owns the infrastructure those models run on, who captures the revenue from them, and who gets to set the terms. This week, both tracks moved fast in ways that are hard to fully process.

Let’s see what next week has in store.