Two stories today that rhyme in an interesting way. GPT-5.6 Luna just got 80% cheaper, permanently. And Guillermo Rauch is arguing that your future customers won't be people who take sales calls. They'll be agents that just start using your product. Put those together and you get a world where AI gets cheaper as it gets more autonomous. The economics of the next wave are clicking into place faster than most roadmaps assumed.
01
GPT-5.6 Luna's 80% price cut is permanent
Thibault Sottiaux, who works on the OpenAI side of this, pushed back on a wave of confusion: the 80% price reduction on GPT-5.6 Luna is not a promotional window. It reflects real efficiency gains in how the model runs, and those don't get walked back. ---
Why it matters: Every budget assumption your engineering team made six months ago about AI API costs is now wrong in your favor. If you shelved a product idea because inference was too expensive to make the unit economics work, it's worth running those numbers again.
"ALG" is the new PLG, and it should make your sales team nervous
Vercel CEO Guillermo Rauch coined a term worth writing down: ALG, agent-led growth. The argument is that the "product-led growth" playbook (let users try before you sell) is being superseded by agents adopting tools autonomously, before any human ever takes a demo call. If your onboarding requires a human to click through it, you're already behind. ---
Why it matters: If your startup's go-to-market assumes a human champion inside a customer account, you need a second plan. The procurement decision for a lot of dev tools is about to be made by a code agent evaluating APIs at 2am, not a VP of Engineering in a quarterly review.
The personal agent that doesn't care which model you're using
Peter Yang, a product leader at Replit, shared his six takeaways from Karan Malhotra, co-founder of Nous Research, on how to actually use Hermes Agent well. The most counterintuitive point: the model underneath barely matters. Because Hermes accumulates memory about you and builds skills around your workflow over time, Malhotra says he barely notices when the underlying model swaps. He also recommends using the /personality command to switch response styles depending on what you need, "teacher" for learning, "concise" for getting answers fast. ---
Why it matters: Yesterday's digest covered Hermes Curator, the system that cleans up an agent's accumulated memory and skills. Today's piece is the flip side: the memory and skills are so valuable that users stop caring about which frontier model is powering the agent. If that pattern holds broadly, the model makers are building commodity infrastructure for whoever wins the agent layer on top.
Shawn Wang (known as Swyx), one of the more reliable barometers of builder sentiment, posted a single line: "lol what are we even doing here anymore guys" alongside a link. No context in the post itself. The engagement (1,094 likes) suggests it landed. The content of the linked image isn't visible here, but the reaction pattern reads like the kind of thing that makes developers laugh because it's too accurate to argue with. ---
Quick hit: Amanda Askell on the difficulty of acquiring a hereditary peerage by force
Anthropic researcher Amanda Askell notes that the UK has really let the standards slip on peerage acquisition. "They just get mad at you if you try." No AI content. Full marks for the observation.