I Was Going to Pick One $200 AI Plan. Both $100 Plans Are Better.
Feb–Mar 2026 I was paying $200/month for OpenAI's Pro plan. Mar–May I swapped to $200/month for Anthropic's Max 20x. This month I was going to sit down with the benchmarks and pick one to commit to.
Instead I'm running $100 on both, paying the same $200 total I was already paying, and it's the most productive AI setup I've had.
Why I kept swapping the $200 plan each month
New models ship constantly and the benchmark leaderboard reshuffles with every release. GPT-5.5 lands on top of SWE-bench. A month later Opus 4.7 takes it back. The smart-looking move was to follow whichever was on top, so I'd cancel one plan, renew the other, and consider myself rational.
That worked for the first two swaps. It didn't work after that.
Benchmarks don't capture how I actually work
Most of my work isn't single-turn model output. It's the Claude and Codex apps, both desktop and CLI, driving real things: Mac app control, homelab maintenance over SSH, coding sessions, and /goal runs left going overnight. SWE-bench doesn't measure any of that.
What it especially doesn't measure is whether the model keeps stopping to ask before doing the obvious next step. That's the number I actually care about.
Claude doesn't spin its wheels
The most concrete difference between the two for me is that Codex keeps wanting input before it acts and Claude just goes. Same task, same prompt, same surface. Codex will plan, ask if the plan is right, ask if it should proceed, ask if it should write tests, ask if it should run them. Claude reads the task, picks an approach, and starts.
On Mac app control specifically, Claude is also noticeably better at actually driving the apps. The mouse goes where it should go. Buttons get clicked. Forms get filled. Codex flails more on the same tasks.
That's most of why Claude felt sharper when I switched in March. It wasn't a model-quality gap. It was an interaction-style gap that adds up over a day of work.
It's not better, it's different
When I switched in March, Claude felt great and I assumed I'd picked the winner. I had not.
Codex iterates more patiently when iteration is what you want. It has more rate-limit headroom in the same price tier. And when I run /goal on Codex CLI overnight, it's happy to grind on a goal for hours, which is the work I least want to babysit. I wrote a longer post on why /goal is the right shape for that kind of run.
Different models, different shapes. Picking one means leaving half the strength on the table.
OpenAI's $100 Pro tier made the swap obsolete
OpenAI launched a $100 ChatGPT Pro tier on April 9, 2026. It slots between Plus ($20) and Pro ($200), gives you the same models as Pro $200, and runs at 5x Plus limits. The Codex sub-quota is 10x Plus through end of May 2026 as a launch promo.
It's aimed straight at Claude Max 5x, which has been $100 at the same 5x level for over a year. That's the missing piece. $100 on each side is the same $200 I was already spending, with both available at once.
How I split the work now
Claude desktop is the default for Mac app control, infra work, and most coding. When I'm at the machine and want something done, Claude gets the first shot.
Codex picks up the slack two ways. The desktop app handles code work when Claude Max 5x's weekly cap fires. /goal on Codex CLI runs overnight when I want to throw something at the wall and see what it built while I slept.
ChatGPT chat handles research, planning, and second opinions. That tab is also where I burn the Codex sub-quota when the desktop session isn't the right surface.
Caveat: the math depends on the limits holding
The $100 Pro tier launched with the 10x Codex promotion explicitly through end of May 2026. If OpenAI quietly drops it back to 5x in June, or tightens the Plus-multiple math more broadly, the split gets less compelling.
For now it's the cheapest way to keep the strengths of both at the same time. If your work crosses agentic coding and Mac app control too, it's worth trying for a month before locking in on a $200 plan.