Issue 60 — 2026-07-11
OpenAI launches GPT-5.6 and the Solana tokens — Terra and Luna — on apps and API
OpenAI has shipped GPT-5.6, the latest iteration of its flagship model, alongside two new tokens — Terra and Luna — that live on Solana. The tokens are available through OpenAI's apps and the API, so any dev can grab them without hunting across exchanges.
Why this matters for us: the bridge between OpenAI's models and the Solana economy is getting thinner — the next batch of tools will pay, settle, and transact without us routing through a dozen intermediaries.
Ternus on Apple's slippery slope
Nick Ternus wrote a sharp piece on Daring Fireball about Apple's direction. Not a rant — a measured observation about how the company's path is tilting, and how hard it is for people watching from the outside to see it happen.
What he's tracking is the slow drift: features that used to serve the power user getting quietly deprioritized, design decisions that look fine until you notice the pattern across a dozen releases. The kind of change that's invisible in a single WWDC keynote but obvious six months later when you try to do something the old way and it just… doesn't work anymore.
This isn't about Apple going bad. It's about Apple going somewhere specific, and the people who love the current Apple — the tinkerers, the devs, the ones who actually use the tools — aren't the ones steering the ship.
Why this matters for us: the tools we use to build things and run the bodega are Apple's. When Apple drifts, we drift with it. Worth paying attention before the shift is a fait accompli.
La pelea no es de ego — es de espacio en la mesa.
— theatlantic.com
#openai-y-anthropic-se-pelearon-la-mesa-y-la-gente-gano-7cb92eSpaceX AI drops Grok 4.5 — a model that actually thinks
SpaceX AI released Grok 4.5 on July 8, 2026. Elon calls it an Opus-class model. That's not the marketing line you hear every week — it means the model has reasoning, planning, and long-horizon task ability in the same league as the best OpenAI and Anthropic offerings. It's built on the same foundation as their earlier Grok 4 line but with enough capacity and training to hold its own against the frontier pack.
What's interesting about this one is the timing and the bet.
Tech workers are tired, and the mood is turning
Lenny's latest survey shows the mood has shifted. Layoffs, burnout, and the AI scramble have left a lot of people feeling like the party's over. The headlines keep telling us the industry is booming, but the people on the ground — the ones doing the actual work — are telling a different story.
The numbers back it up: more folks are second-guessing their roles, more are watching the door. The anxiety isn't just about jobs — it's about relevance. Will what you know still matter in a year? Will AI eat your lunch? Nobody knows for sure, and that uncertainty is what's making everyone tense.
Data is the only moat left — and why the big labs are panicking about it
The Atlantic ran a piece on whether OpenAI and Anthropic are actually different or just running the same playbook. The answer, according to the TLDR Product writer, is that they're converging on the same model architecture and the same training recipes — which means the real competitive advantage is shifting from who has the biggest model to who has the best data.
Data is hard to copy. A model you can download can be rebuilt by anyone with a GPU. The recipes that make it good — the GRPO, the carry, the diversity stacking — those are the real moat. They're in the weights, not in the papers. And once a company has the data pipeline, the training loops, the proprietary datasets, and the people who know how to read the loss curves, that's a wall that compounds. A wall around an empty lot isn't a moat — but a wall around a good data advantage is.
The piece is worth reading because it reframes the whole AI race in plain terms. The big labs aren't racing toward a ceiling; they're racing to lock in the data that makes the ceiling higher. The ones that figured this out first — and kept the recipe secret — will be the ones that matter.
Why this matters for us: the lesson is the same for any company building something real — it's not the brand or the marketing that holds, it's the stuff you can't just copy-paste from a blog post.
How to price tokens for LLMs — a useful framework
Ben Evans lays out a way to think about token pricing that actually tracks what the business does. The key move: separate the cost of running tokens from the value they create, then price the gap. Most models charge by the token, but the token is just a unit of measure — the real question is whether the model earns back its compute cost plus a margin on the work it does.
He breaks it into three buckets. First, the raw compute cost per token, which drops as models get better. Second, the value bucket — how much the token is worth when it solves a real problem for a customer, which is where the margin lives.