OpenAI Codex release notes: 0.144.6 corrects GPT-5.6 context windows

Codex CLI 0.144.6 is a narrow but important correction for teams working with large codebases or long agent sessions. OpenAI refreshed the bundled instructions for GPT-5.6 Sol, Terra, and Luna and corrected their context windows to 272,000 tokens. The stable CLI now has the right model metadata. This is not a launch of three new models or a new operating mode.

Codex CLI is OpenAI's local coding agent for the terminal. A context window is the amount of instructions, code, and conversation history a model can handle in one run.

OpenAI Codex release notes: what changed in CLI 0.144.6

OpenAI published Codex CLI 0.144.6 on July 18, 2026, with two bounded changes:

• The bundled instructions for GPT-5.6 Sol, Terra, and Luna were refreshed.
• The context and maximum context windows for all three models were corrected to 272,000 tokens.

Human step: Install the verified stable version with the command from OpenAI's changelog:

npm install -g @openai/[email protected]

Source: OpenAI Codex changelog for CLI 0.144.6

The correction is deliberately narrow

The first change brought a wider model-catalog refresh to the stable 0.144 branch. OpenAI then narrowed the package before release. The final version keeps the new base instructions, instruction templates, and corrected context metadata for Sol, Terra, and Luna while restoring the other catalog values.

OpenAI explicitly says model visibility and upgrade behavior remain unchanged. Compared with the previous stable baseline, exactly twelve values differ: four fields for each of the three models.

Sources: PR #33972, refreshed model metadata and PR #34009, narrow 0.144 hotfix

Why 272,000 tokens matters in practice

Correct context metadata gives Codex CLI accurate limits to plan a run against. That matters when the agent needs to read many files, follow dependencies through several layers, or keep a longer task coherent.

The 272,000 figure does not mean Codex automatically reads an entire repository or that every item fits in every response. The release note does not promise better code quality either. Test it on work where you already know the architecture, then check which files the agent actually cites.

Source: OpenAI Codex 0.144.6 release on GitHub

Short example: use the new Codex feature

Run this prompt after a human has updated Codex CLI and opened the correct working directory:

Map the architecture of [workflow] in this codebase. Prioritize entry points, data flow, tests, and configuration. Cite files and lines, list important areas you did not inspect, and stop before editing files or running commands.

Good output should:

• Cite real files and lines.
• Separate inspected areas from work that remains.
• Leave the codebase unchanged.

What teams should test now

Choose an existing workflow that crosses several files and compare Codex's map with the view of a technical owner. This shows whether the corrected context information produces a more coherent working brief without confusing a larger context limit with complete understanding.

When the test becomes a recurring workflow, Tool Forge can help connect Codex to the right code and documentation with scoped permissions, secrets held in environment variables or a secret manager, approval before writes, and a retained run log.