# $100 on Codex vs. $100 on Claude

After a month, 40+ automations, and a switch from Claude to Codex, my ideal setup is two $100 subscriptions. Codex gives me more capacity and stronger execution, while Claude is still more enjoyable for creative work.

Author: Maxim Tolmachev
Published: 2026-08-04
Language: en
Canonical: https://maximtolmachev.com/blog/100-on-codex-vs-100-on-claude/
Original source: https://t.me/mtolm/94

![$100 on Codex vs. $100 on Claude](/blog-images/tg-94.jpg)

A month of work, 40+ automations, and my final setup

Over the past month, I ran an experiment: I moved all my automations and my work with Claude (Opus 4.8) to Codex (GPT-5.5 → Sol 5.6).

Migration in a few hours

I did it in literally a few hours, despite having 10 separate projects and 22 ongoing chats. The switch may have been easy partly because of shared memory, shared skills, and a shared conversation database in Retain.

22 ongoing project chats in Codex. Some names are hidden.

Some of my 40+ automations. Yes, two tasks in the screenshot are currently broken 😅

FYI: I use all models with effort: high. This is less a comparison of the harnesses and more about the overall experience and current models.

First, GPT-5.5

At first, GPT-5.5 made me really happy because it did exactly what I asked, usually in one shot — with a single prompt.

However, on large tasks where it needs to generate a lot of content and keep everything in memory, the quality noticeably degrades. You definitely need to check the result with at least one additional pass.

GPT-5.5 is definitely worse at design than Opus 4.8, so presentations and websites became more painful.

Switching to Sol 5.6

I was thinking about going back to Claude because they extended access to Fable 5 and the design quality is solid. But then OpenAI released the new Luna, Terra, and Sol family, and I started testing it actively.

I integrated Luna and Terra into products where volume matters and speed and cost savings are important. I made Sol 5.6, the strongest model, the default for my 22 pinned chats.

And that’s when it got really good: I felt a jump in response quality, design, and performance on complex, long-running tasks. Overall, my quality concerns disappeared. It still sometimes dips on large tasks, but the result is noticeably better than GPT-5.5.

What I liked

Codex has convenient built-in Russian dictation: it’s fast and doesn’t put any load on local hardware. I use it all the time. Claude only has English dictation. I’ve already written about how much voice speeds up my work.

What annoys me

I was doing product fuzzing tasks for a client, and Sol regularly refused to work because it considered them hacking. OpenAI itself says that additional cyber checks sometimes affect legitimate defensive work. For this authorized task, I switched to GPT-5.5.

I also remotely asked computer use to log into my OpenAI account and reset my limit. This option appears in my account once every two weeks. Sol refused, while GPT-5.5 helped. The reset expires if you don’t use it, so I didn’t want to wait until I got home.

Two available full resets in my account.

Codex isn’t as friendly and doesn’t feel as much like “your” assistant: it mirrors you less effectively than Claude. For creative work, when you want to make something cool, Claude is much more enjoyable to talk to. For cold-blooded execution, Codex is still the best option.

My final setup

The main thing I love about Codex is the limit. It feels several times higher, especially with resets. With my $100 OpenAI subscription, I can do roughly 2.5 times more than with an Anthropic subscription at the same price.

Another interesting shift: I used Claude through the CLI in the CMUX terminal, while I use Codex through the official app.

I’ve already written that I spend $180–250 a month on AI. After this experiment, my ideal setup is two $100 subscriptions: one for OpenAI and one for Anthropic.

Decide for yourself which one makes more sense for you right now.
