Four AI subscriptions, and what each one is for

• Claude Max 20x, $200 a month, billed monthly only. Claude Pro is $20 and Max 5x is $100.1
• ChatGPT Pro 20x, $200 a month. OpenAI paused new sign-ups and upgrades to this tier on 10 September 2026. Pro 5x is $100 and Plus is $20.2
• Cursor Pro, $20 a month. Pro+ is $60 and Ultra is $200.3
• Google AI Pro, listed at €21.99 a month. I'm on an offer Google made me for €1 a month, which isn't on the public pages.4

While I was building ChampChase I used Claude Code and nothing else, on the Max 20x plan. Opus did essentially all of that development, up to version 4.8.

Opus 5 didn't work for me. I tried many variations of guardrails and instructions, and every time I gave it another chance it disappointed me within ten minutes. It would agree it had got something wrong, explain why, and then do it again. At that point the only model I found usable for my work was Fable, and it was good.

Fable's problem was quota. On Max plans, Fable models can use "up to 50% of your weekly usage limits," and everything else draws from the same limit.1 On my hardest development days, Fable's half lasted a couple of days at best. After that I still had half the plan left and nothing I wanted to spend it on.

So I moved to Codex. GPT 5.6 at its higher reasoning levels was very good, and Astra, when it arrived, was better than Sol. I used Codex a lot, and I like its harness, especially the experimental feature toggles such as the context summary.

Then Opus 5.5 came out, and it's far better than Opus 5. Right now essentially my whole stack runs on Claude Code.

Claude (Max 20x): the main work. In my experience Claude's models, Fable and now Opus 5.5, do better when the instructions are ambiguous. That means work where the model has to stay on a course I've set, keep the bigger picture in mind and make decisions on the way from A to B without making silly mistakes. I don't currently have a case where a GPT model would do my work better, faster or closer to what I had in my head.

Codex (ChatGPT Pro): the last gate. GPT models are very good when the task is bounded and has a concrete finish line. Today I use Codex for one job: the last check before anything is final, deployed or pushed. I run Astra as an aggressive audit looking for holes, and it finds edge cases a Claude model would overlook. On 24 September it read the plan for replacing this site's articles and caught two things the plan had missed, including an index page that still pointed at the four articles being deleted.

Staying in one harness is part of why Claude does the main work. Its sub-agents talk to each other quickly because they're built for it, and Claude Code can watch something in the background and react when it changes.5 Running two harnesses means setting up another messaging layer between them.

Cursor ($20): the bridge. When both main quotas are empty, I won't pay API prices for flagship models, so I bought Cursor's $20 plan instead. The usage on its Grok and Composer models lasts a long time for the price. They're not as good as Fable, Opus or Astra, but they're good. In Cursor, Grok 4.7 now runs with up to 500K tokens of context, twice its standard 256K.3

I also use Grok through Cursor when I want an opinion from outside the two harnesses I rely on. When I ask Claude about Codex, or about Claude, I find it leans towards Claude. A third provider has no side to take, and Grok gives me the answer straight instead of trying to smooth it over.

Gemini: animating stills. I only use it to turn still images into short scene animations, which needs a Google AI plan.4 At €1 a month it's worth it for that.

I don't like Claude's five-hour limits. For a long time I also believed Max 20x meant twenty times Pro's limits across the board. When I sat down and read the terms, Anthropic's own wording turned out to be "20x the usage of Pro per 5-hour session," and Max plans "also have a weekly usage limit that applies across all models."1 Anthropic doesn't publish how that weekly limit scales between plans. I found the name misleading until I read that.

OpenAI describes its top tier as "20x usage than Plus," and Codex limits also run on a five-hour and a weekly window.2 Neither company publishes its weekly numbers.

This month I went through two or three Cursor subscriptions. Since Opus 5.5 came out I was running through both Claude and Codex too, mostly because I wasn't working efficiently yet. I've found my rhythm now.

Paying per token for flagship models is a bad deal next to a subscription, and I'd rather not go down that route. Per million tokens, Opus 5.5 lists at $4 in and $20 out, Fable 5.1 at $10 and $50, and GPT-6 Astra at $10 and $50.1,2 My own lab records what each Claude Code run would have cost at API prices: between $0.10 and $2.11 per small coding task, depending on the model and the effort level. That's Claude Code's own estimate, not a charge.

I'll keep Claude Max 20x. I may drop Codex to the 5x tier, depending on what OpenAI releases next. With new upgrades to the 20x tier paused, dropping it may mean not getting it back for a while. I probably won't renew Cursor. This changes with every release, and not keeping up means getting worse results than the ones on offer.

• How the plans compare in general. This is one person's work, and different work gets different results.
• That the prices and limits are still current. They change often; these were read on 26 September 2026.
• Anything measured about the models against each other. Nothing here is a benchmark.