CodexZero

Less input. Same task score.

CodexZero saves more tokens.

CodexZero recorded the largest token reduction in our completed comparisons. In the strongest repeated test, it used 15% fewer tokens and matched Codex’s task score.

Install CodexZero See the result
  • Free and open source
  • Installs beside Codex
  • Full results stay available
Highest recorded token reduction for each setup: CodexZero Max used 40 percent fewer tokens, CodexZero Safe 26 percent fewer, RTK 8 percent fewer, regular Codex is the standard baseline, Caveman used 2 percent more, and RTK plus Caveman used 13 percent more.
Highest recorded reduction for each setup across completed benchmark runs. Benchmark summary
15%fewer tokens
Sametask score
3.89Mtokens avoided

Repeated benchmark

Less to process. No score lost.

Codex and CodexZero each completed the same 12 software tasks three times. CodexZero matched the final score while processing 3.89 million fewer tokens.

Regular Codex processed 26.58 million tokens and CodexZero processed 22.69 million while both scored 29 out of 36. CodexZero Max is projected at 19.62 million tokens using the measured Max-to-Standard ratio from the controlled comparison.
12 tasks · three runs each · 36 outcomes per setup
Cache hit ratios across every setup: regular Codex 89.8 percent, CodexZero 90.7 percent, CodexZero Max 91.6 percent, RTK 90.5 percent, Caveman 92.9 percent, and RTK plus Caveman 91.4 percent.
Terminal-Bench mini-panel · cached input tokens / total input tokens

Complete comparison

Savings only count when the task passes.

CodexZero Standard and Max Savings both completed the comparison with fewer tokens than stock Codex. Failed runs, different setups, and smaller samples are marked in the chart.

Comparison of token use, task result, cached input, and estimated cost across CodexZero and other Codex setups. CodexZero Standard used 17.1 percent fewer tokens and Max Savings used 13.7 percent fewer; both passed.
Standard: 17.1% fewer · Max Savings: 13.7% fewer · both passed

Install

Install. Open. Get to work.

Download a release View Windows script View macOS script

Windows desktop. macOS terminal.

After installation

Open CodexZero.

Your existing Codex chats, skills and settings are ready.

How it works

Shorter results when they actually save tokens.

CodexZero keeps the original command result available, removes repeated text when it can, and otherwise leaves the result alone.

01

Run normally

Your commands, model, tools, and project instructions continue to work as before.

02

Remove repetition

Repeated command output can be replaced with a shorter reading copy.

03

Keep the smaller result

If nothing gets smaller, CodexZero sends the original result unchanged.

Three settings

Start with Standard.

Standard uses lean instructions with Codex’s direct tools. Focused adds batching for independent tool calls. Safe keeps the stock model instructions.

Default

Standard

The fastest profile in the paired quick benchmark.

Optional

Focused

Batches independent tool calls for tool-heavy work.

Rollback

Safe

Keeps Codex’s stock model instructions with CodexZero’s output handling.

Quick answers

What you need to know.

Does CodexZero replace Codex?

No. It installs beside Codex, and codex-zero stock opens your regular setup.

What does it change?

It shortens eligible command results before the model reads them again. The full result stays available.

What does Focused do?

Focused can run independent tool calls together. Standard keeps tools direct and is the default.

Will it lower my bill?

CodexZero reduces the input tokens it changes. Your actual bill or plan usage also depends on output, caching, pricing, and your workload.

Ready when you are

Use fewer tokens on your next Codex run.

Install CodexZero