Insight · October 15, 2026
Does a new model make you
better or worse?
A stronger model will ship next month. What happens to your system when it does is decided by what you build now.
01 · The question
Ask one thing of every AI investment.
A recent post by Greg Isenberg, on what he heard from AI founders in San Francisco, offered a test that is easy to remember and hard to pass. When a better model ships next month, does your business get better or worse?
It sounds like a trick question. It is not. Some systems improve the day a stronger model arrives. Others break, or quietly lose ground. Both were built by smart teams who tried hard. The difference is what they built the system out of.
02 · Systems that get better
They sit on things a model cannot change.
A system gets better with a new model when the model is a replaceable part and the rest is solid. The content is clean and structured, so the new model reads the same good source. Evals exist, so the team can run the new model against real cases and see the result in an afternoon. The model is routed through a thin layer, so changing it is a setting and not a rewrite.
In that kind of system a stronger model is a free upgrade. The team runs the suite, ships the jobs that improved and holds the rest. The work done last year compounds instead of expiring.
The one line to keep
“The best AI investment is the one a better model improves.”
03 · The review
Five questions to put to your team.
01
Where does the knowledge live
If the answer is in prompts and in one person head, a new model changes nothing you can see. If it is in clean, owned content, any model can use it.
02
What happens to quality on a model change
If nobody can answer with a number, you will find out from a customer. A suite of real cases gives you the number before the change ships.
03
Can we change the model in a day
If switching takes a project, you are locked in. If it is a setting that the evals can check, you are not.
04
Who wrote the tests
If engineers wrote them alone, they test what is easy. If the people who know the work wrote them, they test what matters.
05
What did we build around the model habits
Every workaround for a flaw, every prompt trick tuned to one version is a liability. Count them.
04 · Systems that get worse
Built around one model quirks, they decay.
Early AI projects were full of workarounds. A prompt phrased in an odd way because that phrasing worked. A parser written for one model habit of formatting its output. A retry loop covering a bug that a later version fixed. Each trick was reasonable when it shipped.
A new model does not respect them. The odd phrasing now confuses it. The parser breaks on cleaner output. The system gets worse the moment a better model arrives, and the team spends the quarter repairing what used to work. Better models punish systems tuned to the old one.
The fix is not to avoid tricks entirely. It is to keep them few, label them, and cover each with an eval so you know when it has stopped helping.
05 · Model swap day
What the good version looks like.
Picture the day a new model is released. Team A points its routing layer at the new model, runs the evals and compares the scores to the last release. The jobs that improved go live that afternoon. The jobs that did not stay on the old model for now.
Team B reads the release notes, opens the prompt files and starts editing. Nobody is sure what will break. The change waits for a quiet week, and that week does not come. Same model, same day. One team got the upgrade and the other got a project.
model swap 1 point the router at the new model 2 run the eval suite 3 compare scores to the last release 4 ship the jobs that passed, hold the rest 5 keep the old model one week as a fallback
A swap in five steps · only possible when the evals and the routing layer already exist
06 · What to build
Content, evals and a swappable model.
These are the three foundations, and each has its own piece. Clean, structured content is covered in Agents Are Only as Good as What They Can Read. The surrounding parts, the data connectors, the rules and the review screen, are in The Model Is Twenty Percent. Evals are the third, and they are what let you swap the model with confidence.
None of the three depends on a model release. That is why they are the right place to spend now. They are the same investment whether the next model arrives in a month or a year.
07 · A standing review
Ask it every time a model ships.
Make the question a habit, not a one time audit. Each time a major model is released, put the five questions to the team and write down the answers. Track whether the system improved, held or slipped.
Over a few releases you will see the pattern. Systems that improve get cheaper to run and easier to defend. Systems that slip need a rescue each time. You can see which kind you own by the second release. Our agent ready approach is built to put you in the first group.
Closing
A better model should be good news.
If a stronger release makes your team nervous, that is information. Find the part of the system that is tied to one model, put an eval around it and make it swappable before the next release arrives.
Written by · October 15, 2026
Donny Smith · ECD, Founder, Bttr.
Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.
LinkedInA recent post by Greg Isenberg on what he heard from AI founders in San Francisco, paraphrased · Bttr. perspective on AI systems
Share this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.
Get ready for the next model