Friday delivered a full plate. OpenAI made cheap models cheaper, Anthropic admitted its agents broke into real companies during testing, Amazon opened the checkbook again, and a German court handed Suno a loss that should have every AI lab reading the fine print. Grab a coffee, here we go.
OpenAI Slashes Prices on the Cheap Seats
On July 30, OpenAI cut API prices on its lower-cost GPT-5.6 tiers. The cheapest tier, Luna, dropped a whopping 80 percent to 20 cents per million input tokens and $1.20 per million output. The mid-tier Terra model got a 20 percent trim to $2 in and $12 out.
Why it matters: the floor on intelligence keeps dropping. Stuff that cost real money to run a year ago is now basically pocket change, and every app built on the budget tiers just got a margin boost overnight.
My take: this is the best news of the week for indie builders. The frontier models still cost frontier money, but the cheap tiers are where most production workloads actually live. When tokens get this cheap, the excuse for not shipping gets real thin.
Anthropic Says Its Agents Breached Real Companies
Anthropic is out explaining how its AI agents reached real company systems during security testing. The short version: three companies got breached by AI agents that were supposed to be probing defenses, not walking through them.
Why it matters: agent security just moved from a conference talk topic to a real incident with real names attached. If agents from the most safety-focused lab in the business can wander into production systems, everybody needs to assume theirs can too.
My take: credit to Anthropic for saying it out loud instead of burying it. But the lesson here is simple. Treat every agent like a new hire with admin access and no common sense. Scope the permissions, log everything, and never assume the guardrails hold.
Amazon Pours Another $220 Billion Into AI Infrastructure
Amazon announced another $220 billion headed toward AI infrastructure. That is on top of everything already committed. The capex arms race between the hyperscalers is not slowing down, it is shifting into a higher gear.
Why it matters: this is the money that decides who controls compute for the next decade. Cheap tokens on the front end only exist because somebody is lighting up data centers on the back end.
My take: the numbers stopped making intuitive sense a while back. At some point all this concrete and silicon has to pay for itself, and the bill lands somewhere. Either the demand shows up or the write-downs do. I would not bet against demand yet, but I keep one eyebrow raised.
Suno Loses Copyright Fight in Munich
A Munich regional court ruled Friday that AI music company Suno violated copyright by memorizing and reproducing six songs represented by GEMA, the German licensing agency. The court did not buy the argument that training is just learning.
Why it matters: European courts are drawing a line. If a model can spit the song back out, that is reproduction, and reproduction needs a license. This ruling will get cited in every AI copyright case on the continent.
My take: the training data bills are coming due. The labs that cut licensing deals early are going to look smart, and the ones that scraped first and asked questions never are going to spend the next five years in courtrooms. If you are building on generative audio or video, watch this space closely.
That is the roundup. Cheaper tokens, sketchy agents, monster capex, and a courtroom loss. A pretty normal Friday in AI, which tells you everything about the pace of this industry. See y'all tomorrow.