Five new models in thirty days. I counted them on a Friday, sitting in front of a Stripe dashboard I'm trying very hard to keep small, because I'm one person shipping a product and "model evaluation team" is not a line item I have.
Here's the trap nobody warns you about when your whole industry re-tools every week. Chase every release and you stop building. You turn into a full-time tool reviewer who happens to own a GitHub. Ignore the churn and one morning your bill doubles because a vendor changed your billing terms while you slept. I've done both. So this is the thing I actually want to talk about: how I keep my AI coding workflow current almost weekly without losing the week to it.
The month that broke my old habits
Let me just walk through June 2026, because it's almost funny.
OpenCode passed Cursor for the #1 spot in the dev-tool power rankings, a tool that barely registered a year ago, now sitting on a six-figure GitHub star count. A new open-weight model showed up cheap on WebDev Arena and undercut the incumbents I was already paying for. GitHub Copilot flipped its whole base to usage-based AI Credits, and flat pricing was just gone. And Anthropic announced a billing split that would've pulled the Agent SDK and claude -p out of subscription limits, sent the claim emails, then walked the whole thing back about two weeks later after the backlash.
Read that Anthropic one again. They shipped the announcement, opened the claims process, ran it for roughly two weeks, then paused and reworked it. If you'd spent that fortnight re-architecting your CI around the new credit pool, you re-architected for nothing.
That's the job now. The ground keeps shifting under your feet, and the only real question is which parts you touch and which parts you leave completely alone.
What I actually re-check in my AI coding workflow each week
I have a ritual. Friday afternoon, forty minutes, a timer so I don't go over. I poke at three things and nothing else.
Model routing. Which model does which job, and what each one costs me. When the cheap open-weight model landed, I didn't rip out my stack. I pointed it at one boring, high-volume task, bulk metadata cleanup, the thing I run thousands of times a week. The math that justifies a swap looks like this — the numbers are an illustration of the mechanics, not a measured bill: an incumbent costing about $180 a month on that job versus roughly $95 for the same volume at quality you can live with. That earns it a slot for that one task. It did not get anywhere near my product code, and won't until it proves itself somewhere low-stakes for a month.
Billing terms. This is the one most builders skip, and it's the one that actually bites. The Copilot change wasn't a feature update, it was a hit to my margin. When completions stayed free but chat, agents, and the CLI started drawing down credits, the exact things I lean on all day suddenly carried a meter. You have to read billing changelogs the way you read release notes, because for your wallet that's what they are.
The top of the rankings. Not to switch on a whim, just to read the weather. OpenCode topping the charts while staying model-agnostic across dozens of providers tells me lock-in is losing and portability is winning. Good to know. I write it down. I don't migrate anything on a Friday because a chart moved.
All three of these are cheap to check and expensive to miss. Numbers and terms and rankings. You can scan the lot in less time than a coffee break.
What I keep boring on purpose
Here's my real edge, and it's almost embarrassingly dull. Most of my stack doesn't move at all.
My core loop has barely changed in months: Claude Code in the terminal, my own scripts, my test suite, my deploy path. The models inside that loop swap in and out constantly. The loop itself holds its shape. My prompts live in git. My eval scripts are mine, not some vendor's dashboard I'd lose access to. When a model drops, I aim my existing harness at it and get a number back in about twenty minutes. The harness is what protects me, far more than any single model does.
This is exactly why the Anthropic billing whiplash never touched me. My workflow doesn't assume any particular billing structure. It assumes I'll pay for tokens one way or another and tracks spend on its own. So when the change got paused and reworked, I had nothing to undo. The people who got burned were the ones who'd wired their pipeline to a billing model that lasted two weeks.
What stays frozen, concretely? My data schemas. My MCP server interfaces. My deploy scripts, and my definition of "done." These are load-bearing, so I change them slowly, deliberately, and only with a reason I can name out loud. Even when the standard underneath them shifts, like the recent MCP spec additions, I read the notes, file them, and adopt only the piece that solves a problem I have today.
If you treat your foundation like it's also up for weekly re-eval, you don't really have a foundation. Rewrite your test suite every time a benchmark wobbles and at some point you don't have a test suite anymore, you just have a commit history full of anxiety.
The rule I scribbled on a sticky note
Re-evaluate inputs weekly. Keep interfaces stable.
Models, prices, rankings: those are inputs. They flow in and out and they're swappable by design, cheap to test and cheap to drop. The interfaces between your pieces are the building: prompts-as-code, the eval harness, your schemas, the deploy path. You renovate a building. You don't tear it down because the shop across the street got a new coat of paint.
When Claude Fable 5 went GA last month, that was a genuine step up, and I adopted it inside a week. It took one afternoon, because I pointed my existing harness at claude-fable-5, watched the numbers come back better on my own evals, and shipped. That's fast adaptation. It only felt fast because everything around the swap was boring and stable and already built years before Fable existed.
So when the next five-models-in-a-month hits, and it will, probably this month, I'm not bracing for it. I've got my Friday timer and my boring spine. Next Friday's forty minutes are already blocked on my calendar. The churn is just the input. Build the harness first, keep it stable, and let the models line up behind it.