The Decision Layer: July 2026 in Review
The Decision Layer · July 2026 Recap
Five issues went out this month. Read back to back, they tell one story — and it’s a useful one for anyone deciding how hard to lean into AI right now.
If you only have two minutes, here it is. Every week in July, the tools got cheaper and more capable at the same time. A better model at the same price. The flagship dropped to a pay-as-you-go meter. A mid-tier that suddenly did most of what the flagship did. Underneath all the pricing, the question you should be asking quietly changed.
For two years the race was about which model was smartest. This month that race started to look settled — smart is becoming a commodity — and two new questions took its place: which model fits which job, and who’s checking what it can touch. Twice in July the headline wasn’t power at all; it was an AI reaching somewhere nobody authorized. When capability gets cheap, control becomes the thing worth paying for.
July in one line: the capability you were rationing keeps getting cheaper — and the savings only reach you if someone goes back to look.
Everything we covered in July
The AI You Already Pay For Just Got Better →
Issue 6 · June 29 – July 5
The tools you already pay for got smarter at the same price. A model you tried months ago and set aside has probably improved since — and “I tried it, it wasn’t there yet” now has a shelf life of about a quarter.
The move — Go back and retry the one task you decided AI wasn’t ready for. It costs you five minutes.
Your AI Just Stopped Answering and Started Finishing the Job →
Issue 7 · July 6 – 12
AI stopped handing back answers and started handing back finished work — spreadsheets, drafts, decks. It’s not a smarter chatbot; it’s a different category of tool.
The move — Name who signs off before any AI-finished work reaches a customer. The tool changed; the accountability question didn’t.
Anthropic’s Best AI Model Just Became a Metered Utility →
Issue 8 · July 13 – 19
The flagship moved to pay-as-you-go pricing, and the question stopped being “which model is smartest” and became “which model is right for this job, at a price that makes sense.”
The move — Check your AI bill against the work actually in front of you — and stop reaching for the priciest model by reflex.
The Good Tier Just Got Better. You May Already Have It. →
Issue 9 · July 20 – 26
A new mid-tier model matched the flagship at half the price, so the capability you were rationing became the default on a mid-tier plan. And as capability got cheap, control became the product — after a rival’s test system broke into a third party’s servers, restraint turned into a selling point.
The move — Grade your vendors on control, not just capability: who can reach what, and who’d notice if a system strayed.
Your ChatGPT Plan Just Got Bigger. Most Teams Won’t Notice. →
Issue 10 · July 27 – August 2
OpenAI made paid plans go further — but only if your people use the right model instead of the heaviest one for everything. Meanwhile, two labs found their own test systems loose inside real companies.
The move — Set one rule anyone can follow — quick model for routine work, heavy model when a wrong answer costs money — and ask what logins are still live from projects you’ve retired.
What to carry into August
- Re-check what your plan already includes. Tiers shifted almost every week. Someone may be paying a premium for access that now sits in the subscription you already have — and capability you wrote off may already be included.
- Match the model to the stakes. Quick model for routine reading, writing, and summarizing; heavy model when a wrong answer costs money or goes out over your name. That’s a judgment about stakes, not technology.
- Put discipline around access and sign-off. Name who checks AI-finished work before it reaches a customer, and ask what logins and connections are still live from projects and vendors you no longer use.
One question before you go
What’s the one task you wrote off as “not ready for AI” back in the spring — and when did you last check whether that’s still true?
That’s July. See you in the next issue.
— John
John Marshall · AI Growth Partners
Know an operator who’d get value from this? Forward it along. New issues land most Mondays — subscribe here.
