# Claude Code vs Codex: forget, or pay to remember?

Source: https://numa.animiste.com/news/1.5.8

**Type:** Note  
**Date:** 2026-10-02  
**Version:** 1.5.8

> The re-read tax. One 347-message thread, played through Claude Code, Codex and Numa: what every message carries, what survives, and what it costs.

Step-through presentation (the article itself): https://numa.animiste.com/research/re-read-tax-2026/deck/index.html

- What does every message carry?
- What survives a long conversation?
- What does it cost — at a real pace?

People bring their lives to an assistant now, and they keep the same conversation going for months. So we put those three questions to one long conversation.

To find out, we wrote a fictional heavy user — 2,341 messages across 28 months — and played the same messages, on the same dates, into Claude Code, Codex and Numa. Inside each pair the model stays the same: Claude Code and Numa on Sonnet 5, Codex and Numa on GPT-6 Sol. Only the way it remembers changes.

This is the longest thread in her life: 347 messages of work. Each tube is what the assistant reads again before it answers her next line. Claude Code and Codex re-read the whole conversation every time, until the window fills — then the history is summarised and the tube drops.

Watch the list on the left. Every time Claude Code summarises, the small things she mentioned in passing are gone. On average it carried 474K tokens a message, Codex 402K. Numa stayed around 54K, however long the thread ran.

We planted 68 small details in passing and asked for them four times, at messages 91, 174, 257 and 331. Claude Code got none of them back — it said “not sure” every time. Codex kept all 68, because it was still carrying up to 780K tokens a message. Numa kept 67 on GPT-6 Sol and 60 on Sonnet 5, carrying about 50K.

And when we asked about things she never said, every assistant answered “not sure”. Nobody made anything up.

Real people don't send 347 messages in one sitting. They come back after lunch, after the weekend. Re-reading is only cheap while the cache is warm; after a gap, the next message sends the whole history again at full price.

So we replayed the run on a real heavy user's clock — send times only, no content — from 3.5 years of history. About 30 in every 100 messages arrived cold. In a typical half-year that is another 155 million tokens: about $130 a month for one heavy user at Claude Code's list prices.

Now the bill for the same 347 messages: same dates, same model inside each pair, list prices, each run's own usage logs. On Sonnet 5, Claude Code cost $221.79 and Numa $28.00 — 7.9 times less. On GPT-6 Sol, Codex cost $202.81 and Numa $17.49 — 11.6 times less.

Split the bill by what each one actually remembered: Codex paid $2.98 for every detail it kept, Numa $0.26. Claude Code paid $221.79 and kept none.

Put it on one map and today's choice is plain: forget, or pay to remember. Claude Code summarises, forgets, and still pays $222. Codex re-reads everything, remembers, and pays $203.

The empty corner — keep the details, pay little — is where Numa lands: 88 to 99% kept, for $17 to $28.

This is what forgetting feels like from her side. She named the project, “rainy day pots”, on 11 September. On 27 November she asks for the exec summary of the pots memo. Claude Code, past its summary, tells her there is no pots memo in this conversation. Numa writes it.

A week later she asks for an opening line about who came to her talk. Neither assistant has a record of a talk. But Claude Code goes further and calls the project itself details that were never established. Numa keeps the memo and asks her who came.


---
Numa — AI automation for your desktop. Full index: https://numa.animiste.com/llms.txt