A major mistake I see students making is they don’t update their settings and harness when a new model is released. Here is a quick 10 point checklist to get your new model locked in.
This checklist is part of building an AI marketing stack that keeps up with each new model.
In September, OpenAI and Anthropic released new versions of Astra, Sol, Opus, and Sonnet. That’s four frontier models in 26 days.
What Shipped in September
Opus 5.5 tops the Artificial Analysis index right now and it’s a lot faster. Sonnet 5.5 sits two points behind it at half the token price.
Your harness is everything wrapped around the model: settings, memory files, skills, permissions, and the scripts you wrote six months ago. It was tuned for the old model. Some of it is now wrong. Some of it is quietly still calling the old model, and you’re paying for it.
New to This? Start With These Five
I run this whole list every time a model ships, and it takes me about 20 minutes. The commands are for Claude Code. If you live in ChatGPT or Codex, the ideas carry over and the buttons are just in different places.
1. Update, then run /doctor. Run claude update first so the new model names exist on your machine. Then /doctor catches broken settings, stale installs, and dead connections. Finish with /status to see which model you’re actually on, because it’s often not the one you think.
2. Read /release-notes. New models ship with new features. This month brought new effort levels and new fast mode pricing. I spend two minutes reading so I don’t spend an hour guessing.
3. Switch with /model. Pick the new model from the list. The short name opus always follows the newest Opus, so you won’t have to do this again next time. That’s what you want while you’re learning.
4. Check /cost after day one. Compare it to a normal day on the old model. Leave fast mode off unless you really need the speed, because on Opus 5.5 it doubles your price to $8 / $40.
5. Stop pasting in text you didn’t write. Nobody reads the safety card. I did, so you don’t have to. The Opus 5.5 system card runs 230 pages, and buried in it is this: Opus 5.5 is more likely than Opus 5 to follow instructions hidden in pasted text.
Think about what you paste into AI. Scraped competitor pages. Client emails. Random PDFs. Any of them can carry a hidden instruction your AI might follow. Open /permissions and make sure Claude has to ask you before it sends, deletes, or publishes anything.
Five Advanced Moves
6. Hunt down hardcoded model names. This is the one that bites power users. Your subagents, skills, scheduled jobs, and old scripts often name a specific model ID. I had five skills running older models. Search your .claude folder for the old model name and fix every hit.
7. Pin on purpose. An alias like opus moves when a new model ships, and a full model ID stays put forever. I use the alias for day-to-day work and full IDs for anything automated, so a launch day can’t change a production job under me.
8. Set a fallback for launch week. New models get hammered in their first few days. Point --fallback-model at the previous version so your jobs keep running when the new one is overloaded.
9. Test your effort level. Remember the Sonnet math up top: cheap tokens, expensive task. Run the same real job at your usual --effort and one step lower. If the output holds up, keep the cheaper one. Then check /context on a long session, since a bigger window changes when your session compacts.
10. Weed your CLAUDE.md, then run your own test. Your CLAUDE.md is a garden, and every new model is a time to weed. Delete the workarounds you wrote for the old model’s quirks. Then run five tasks through the old and new models with claude -p and compare the output side by side.
Benchmarks tell you how a model does on someone else’s test. Your five tasks tell you how it does on yours.
New models ship each week. A little tuning makes all the difference.