Codex vs Claude Code:
The Honest 2026 Field Report

I use both every week: for this website, for client landing pages, for tracking scripts and for the small browser games in the Game Center. This is not a benchmark article. Benchmarks change every month and rarely match what happens in a real repository on a Tuesday afternoon. It is a field report: what each tool is, what it costs in practice, how the limits really behave, and the prompts and hidden settings that make the difference.

Freshness warning. Plans, prices and limits in this space change often. Everything below reflects what I could see in September 2026. Before you pay, check the official pricing pages linked at the end.

What you are actually comparing

Both names describe a family of surfaces, not a single app. That is the first source of confusion. When someone says “Codex is faster” or “Claude is smarter”, ask which surface and which model they used.

The takeaway: at the surface level they have converged. You get a terminal agent, an editor plug-in and a cloud mode on both sides. The real differences are in the models, how usage is counted, and how each tool behaves when a task gets long.

The models underneath, and where Fable fits

Claude Code lets you switch models with /model. In September 2026 the Claude line-up is, from heaviest to lightest:

Codex runs on OpenAI’s GPT‑5 family, including models tuned specifically for coding, and exposes a reasoning effort setting (low, medium, high and above on some plans). Higher effort thinks longer before acting: better on hard problems, slower and heavier on your allowance.

A practical routing rule I use: plan with the strongest model, implement with a mid-tier one, and check with a fast one. On Claude that is often Fable or Opus in plan mode, then Sonnet for the edits. On Codex it is high effort for the plan, medium for the build.

Usage limits and credits, explained without the marketing

This is the part most people misunderstand, and it decides whether you feel the tool is “generous” or “always out”. Neither product is unlimited, even on the top consumer plans.

Rolling session window

Claude subscription plans count usage in windows of a few hours. Hit the cap and you wait for the window to reset. Codex on ChatGPT plans works in a similar way, with short-term limits per window.

Weekly cap

On top of the short windows, both vendors have added weekly limits for heavy users. Heavy all-day agent use can reach the weekly cap before the week ends.

Credits / extra usage

When you run out, both sides now let you keep going by paying for more rather than upgrading the whole plan: OpenAI sells Codex credits, and Claude plans can switch on extra usage billed separately. Useful for a deadline week, dangerous as a habit.

API billing

Both tools can run on a developer API key instead of a subscription. You pay per token with no window, which is predictable for automation but easy to overspend in an interactive session.

What “uses more”

Bigger model, higher effort, longer conversation, larger files read, more parallel agents. A single hour-long session with a huge context can cost more than twenty short ones.

Which plan makes sense?

Habits that make the same plan last much longer

These are not tricks. They simply stop you paying the model to re-read things it no longer needs.

• Clear between tasks. /clear in Claude Code, a new session in Codex. Every message resends the conversation so far; a stale conversation is a tax on every future reply.
• Compact before you clear. If you want to keep the thread, /compact summarises it. Add a hint: /compact keep the file list and the failing test.
• Point, don’t paste. “Look at src/cart.js line 80” is cheaper and better than pasting 400 lines into the chat.
• Trim logs. Paste the last 30 lines of an error, not the whole build log.
• Keep the memory file short. CLAUDE.md and AGENTS.md are loaded into every session. A 2,000-line memory file is paid for on every message.
• Route by difficulty. Switch to a lighter model for tests, renames and copy edits.
• Plan once, build once. A bad plan executed costs three times: the build, the revert and the rebuild.
• Batch small asks. Five small changes in one clear message beat five separate conversations.
• Stop runaway loops. If the agent tries the same fix twice, interrupt it (Esc) and give it the missing fact yourself.

Prompts that work on both tools

Good agent prompts read like a short ticket written by a senior colleague: the goal, the constraints, how to prove it is done. Copy these and adjust the details.

Read the checkout flow in src/checkout/ and explain how a discount code is applied today.
Then propose a plan to add a "free delivery over OMR 20" rule.
Do not edit any files yet. List the files you would change and the risks.

The contact form sends twice on mobile Safari. Find the cause.
Write a failing test that reproduces it first, then fix it,
then run the full test suite and show me the output.
Keep the change as small as possible.

Add a new blog card to blog/index.html for the post in blog/new-post/.
Copy the structure of the newest existing card exactly:
same classes, same image sizes, same date format. Do not restyle anything.

Review the diff on this branch as a strict senior reviewer.
Only report real bugs, security problems and broken edge cases,
with file and line. Skip style opinions. If it looks fine, say so.

Summarise what we changed today, what is unfinished, and the exact
next step, in under 15 lines. Save it to NOTES.md so a fresh session can continue.

Notice what these have in common: a clear stopping point (“don’t edit yet”, “show me the output”), a scope limit (“as small as possible”, “don’t restyle”), and a way to verify. Vague prompts like “make it better” are where both tools waste the most usage.

Hidden tips most people never find

What to expect, honestly

My own split today: Claude Code for this website’s heavier work and anything involving writing, Codex for quick fixes and a second-opinion review. If you are just starting, pick the one attached to the subscription you already have, use it seriously for two weeks with the habits above, and only then decide whether the other is worth paying for.