OpenAI Drops GPT-5.4 With a Million-Token Window
OpenAI rolled out GPT-5.4. A million tokens of context. Plus it can run multi-step workflows on a real desktop without holding your hand.
Here's the kicker. It scored 75% on OSWorld-V, the benchmark for actual desktop productivity work. The human baseline is 72.4%. Yeah. The model did the office job better than we do.
Why it matters: this is the first time a general purpose model beat human performance on real knowledge work in a public benchmark. Not a contrived test. Real spreadsheet wrangling, real form filling, real email triage.
My take: if you run a small business and you're still doing your own data entry, you got about six months before a competitor automates the same work for fifty bucks a month. Stop reading about AI. Start shipping with it. The folks who win this round won't be the ones with the smartest model. They'll be the ones who actually built something useful and put it in front of customers.
Anthropic Buys Up SpaceX's Memphis Data Center
Anthropic signed a deal to grab all available compute at Colossus 1, the SpaceX facility in Memphis. That's over 300 megawatts coming online for Claude.
This is a direct response to the company posting 80x revenue and usage growth in Q1. Eighty times. Folks have been getting timeouts and rate limits on Claude Pro and Max for months now.
Why it matters: AI is bottlenecked on power and silicon, not ideas. The big labs are now fighting over megawatts the way they used to fight over engineers. Memphis just became one of the most important zip codes in tech.
My take: I said this a year ago and I'll say it again. The next decade of AI gets decided by who can get electricity to a data center. Forget the model wars. Watch the substations. Texas should be building right now, but our regulators are still fighting last year's battles.
Anthropic Ships Ten Agents Aimed at Wall Street
Anthropic also rolled out a set of ten preconfigured agents built for investment banks, asset managers, and insurance companies. These aren't chatbots. They're built to run actual workflows like portfolio rebalancing, claims review, and KYC checks.
Why it matters: financial services has been the slowest sector to adopt AI because of compliance overhead. If Anthropic can get a foothold here, they lock in long contracts at fat margins. And every other vertical is watching what happens.
My take: junior analyst jobs are going away. Not in five years. Now. If you're paying somebody six figures to summarize earnings calls, you're already behind. The good news is the analysts who learn to direct these agents will be worth a whole lot more than the ones being replaced. Pick a side and get moving.
That's the rundown. Stay sharp out there.