Last week, Anthropic unceremoniously lobotomized Claude Code mid-week, making it unusable for countless people who had come to rely on it. On Monday, Claude was a reliable coding partner. By Wednesday, it was a reckless, half-assed clod arguing with itself. Honestly, with how hazy, slow, and confused it got, it felt less like it was processing tokens and more like it was out back tokin'.
It went in circles, pillaged my quota, and set fire to whatever I let it near.
It hit me particularly hard because Claude started destroying my very carefully designed client work and trashed the agent harness I've used to ship hundreds of thousands of lines of production code. To give you an idea of the scale, here's the card for one of my client projects. I routinely run three or four at a time:
Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug
·················▓·▓··░····▒···············░·▓···░▓█
Mon ··················█▒▒·▒█·░░▓·········█·····▓·██··░██
··················░▒░▒░█░░▒▒······▒·▒▒·▒···▒░██··▒██
Wed ···················░░▓░░▒▒▓▓······▓▒▒██▒···▓▓▓▓·░███
·····················░·█·▒▓░·········█·█···░▓▓█░·███
Fri ·················░···█▒·▒·▒▓·······▒···▓··▓·▓█▓▒·▓▒▓
·················░····░···░▒·······█······▓·░░···▓▓▒
Less ░ ▒ ▓ █ More
All time · Last 7 days · Last 30 days
Favorite model: Opus 4.7 Total tokens: 85.1b
Sessions: 420 Longest session: 18d 2h 19m
Active days: 110/246 Longest streak: 21 days
Most active day: May 14 Current streak: 21 days
Input 45.4m · Output 236.8m · Cache read 81.7b · Cache write 3.1b
But let me be clear: I absolutely hate "token-maxxing." Bragging about raw token volume is like looks-maxxing for tech bros. What actually matters in production is ROT-MAXXING: Return on Token. You want maximum business value for every token spent.
When Anthropic pushed this update, my ROT went NEGATIVE. My design and build process relies on agents running for hours at a time. It's usually highly efficient, so hours had gone by before I actually discovered the issue. By then, real damage had been done, and real money—in wasted tokens—had been spent.
I had to stop all client work mid-stream. The tips of all those branches now contain trash from "Clod Code."
The Codex Rescue
Luckily, my harness can be used across Claude Code, Codex, and Antigravity. I hadn't materially built much in Codex for 6 to 9 months because Claude had been so stable and controllable. But in the middle of this disaster, I plugged in Codex and successfully identified, triaged, and fixed my harness and client builds.
With Codex plugging along, I decided to dig a little deeper and kick the tires on Gemini 3.7 Flash. Here are a few direct learnings from that pivot, followed by what I learned about Anthropic's corporate and product "trust."
First, Codex. ChatGPT has made amazing strides with their models. Many months ago, Claude easily overtook Codex in model quality, leading me to switch to ChatGPT mainly for voice mode and ideation. (ChatGPT's voice models and voice chat are peerless, by the way. Get some good headphones and go talk to ChatGPT in the app. It's incredible.)
Now, though, their output quality is at least on par with the smart, hard-working Fable. Its writing style and UI creation actually seem better to me. GPT 5.6 Sol might be my new go-to model. Both Codex and Claude ship with a /btw feature, but Codex's is far better. It opens a multi-turn side chat you can dip into and out of at ease, whereas Claude is a one-message-and-done experience in the main chat UI.
Codex's downside is still the app itself. When Claude Code organizes its work well, it fans out Agent Teams and multiple subagents flawlessly, getting work done quickly. Codex seems far more limited in its ability to deeply manage those subagents and teams. Yes, it can spin them up and put them to work—but not as many, and not hierarchically deep like Claude. However, Codex is VERY THOROUGH.
BTW: As I write this, here is Codex's review of broken Fable project I'm currently fixing:
Verdict: NOT VERIFIED; BOUNDARY REBUILD REQUIRED.
The Gemini 3.7 Flash Experiment
Next, Gemini. In my drive to keep client projects moving, I also checked out Gemini 3.7 Flash, which Google shipped on August 13 as a coding and agent model at $0.75 per million input tokens. This was a funny experience.
For critiquing others' work and helping me think through issues, Flash was great. It is super, super fast—like, nearly instant. I used it to help me plan the re-architecture of my agent harness (which is now built), and together we fleshed out a solid plan.
Then, just for fun, I gave that same plan to both Gemini 3.7 Flash and Codex to see what would happen. Gemini started slinging code and files like mad. The code printer went BRRRR! In its cute, naive mind, it delivered the entire roadmap in about 5 to 7 minutes (maybe 10), including the PR. Lots of code. Many files touched. I was floored, thinking, Either I'm looking at the future of software dev, or this code is humorously wrong.
Codex did the same rewrite—uninterrupted—for about 10 hours. It was meticulous.
The verdict? Gemini's code wasn't even window-dressing. It somehow shipped green tests, but everything else was completely useless. It was pure engineering theater.
Codex got the solution 90% correct. Everything was done well; it just missed a few things and overstepped in another area (changing the scope and purpose of my product documentation).
What I learned about The Flash: It will make a fantastic agent model. With very tight context in well-scoped missions, Gemini is great and cheap. But Codex is what I'll be using for real engineering work. Flash isn't dumb, but it isn't built for these long, multi-turn, highly orchestrated builds. If you need agents adding to a cart, making recipes, or exercising small, tightly scoped judgment? Yes.
A Complete Loss of Trust
Finally, Anthropic.
I've had over a year of amazing results with Anthropic's models. I have shipped so much. I even gracefully navigated the 4.6 to 4.8 nonsense that wrecked so many others. But this week was painful, costly, and eye-opening.
And here's what makes it worse: they already knew.
On April 23, Anthropic published a postmortem on the last six-week Claude Code quality collapse. Three causes, and every one of them was a product decision, not a model regression. Default reasoning effort quietly dropped from high to medium to shave latency. A caching bug wiped prior reasoning out of sessions every turn. A system prompt instruction told the model to shut up in 25 words or less between tool calls. Nobody's model got dumber. Somebody made choices.
They wrote down the fix, too. Per-model evals on every system prompt change to Claude Code. Soak periods, broader eval suites, and gradual rollouts for anything that trades away intelligence. A larger share of internal staff on the exact public build. New tooling to review and audit prompt changes. They even reset usage limits for every subscriber.
That was four months ago. Whatever soak period was supposed to catch this one didn't catch a damn thing.
What Anthropic did to their customers this week is something I won't soon forget. I'm already changing my subscription plans and heading back toward Codex. I also just don't trust Anthropic's product team anymore. The damage they did was startling. Anthropic has wasted my time and my money—neither of which I will get back. Given their days-long silence and milquetoast response, whatever goodwill or trust I had in them is absolutely gone.
Anthropic, if you're reading this (hah, you're not, I have like 10 followers), you owe me and massive swathes of your affected customers an apology, a refund, and a roadmap to ensure your product strategy isn't as braindead as you just made Claude this week. You reset everyone's limits in April. Do it again—and this time, show your work.
Jason Vertrees is the founder of Heavy Chain Engineering, which helps lower middle-market vertical SaaS companies and PE firms turn scattered AI usage into measurable delivery leverage — 85% faster feature velocity, six-to-eight-week projects shipped in days. If you want help building an AI-native engineering organization, book an AI Delivery Assessment or email jason.vertrees@gmail.com.