Why I Cancelled Claude Max After 13 Months and What I’m Testing with Codex Next

I cancelled Claude Max yesterday.

I already wrote about dual-wielding Codex and Claude Code, and then about why the economics shifted enough to cut the $200 plan. This post is the actual decision.

I stayed on Claude Max for 13 months and paid US$1,892.38 over that stretch. It became part of how I shipped almost everything: this site, the course platform, STRATUM, DIALOGUE, the blog pipeline, and the video system. So cancelling it is not just a cost decision. It is a workflow decision.

Why I cancelled

Three things, in order of importance:

1. Codex with GPT-5.4 is closing the gap fast. I routinely pitch the diagnosis and plan from one AI to the other and get them to critique each other's work. Over the past few weeks, Claude Code has been agreeing with Codex's findings more and more — and the plans coming out of Codex are often approved by Claude with only minor changes. That was not the case two months ago.

2. The price difference is hard to justify. Codex is $20/month with generous usage limits. Claude Code Max is $200. When the execution quality converges, 10x pricing has to be earned every single day.

3. I want to find out what happens. Can I maintain my current pace of shipping without the tool I have relied on most heavily for over a year? I genuinely do not know. That is the experiment.

I have to admit — the recent reliability issues at Anthropic did not help. And Opus 4.6 with the 1M context window has felt noticeably slower than GPT-5.4 at equivalent thinking levels. That might be temporary. But it made the decision easier.

If you want the fuller reasoning — the newsletter story, the SSE bug, the uptime numbers — yesterday's post unpacks all of it.

The 30-day test

For the next 30 days, I am running Codex at $20/month as my primary tool. I want to see what actually changes in four areas:

1. Shipping speed — features deployed per week, compared to my March average
2. Code quality and rework — how often I have to revisit and fix something the next day
3. Reliability during longer sessions — how long I can work before the tool gets in the way
4. Avoidance behavior — tasks I defer or route differently than I would have with Claude

That last one matters more than it sounds. A tool does not need to fail dramatically to be expensive. Sometimes it just quietly changes what you are willing to attempt.

The commitment

I do not want this to become one of those build-in-public decisions that sounds dramatic for 24 hours and then quietly disappears.

On May 2, 2026, I will publish a follow-up: what broke first, what got easier, what I missed from Claude, and whether I would make the same decision again. This could end with me resubscribing. Either way, the follow-up will be more useful than the cancellation itself.

I know this is the third post on this topic in three weeks. After this, I shut up and ship for 30 days.

---

That's it from me for now. If you have made a similar switch, I would genuinely be interested in what changed for you after the first few weeks.

Cheers, Chandler