OpenAI's New CLI Is Apache-2.0 and Written in Go, but Its README and Docs Disagree on the Go Version

OpenAI published openai-cli, an official command-line client for its REST API, under the Apache-2.0 licence, and it reached 202 points on Hacker News on Oct. 22. The repository README says it requires Go 1.26 or later to install from source. OpenAI's documentation page for the same tool says Go 1.25.

The repository showed 999 stars, 83 forks, 4 open issues and 23 open pull requests when fetched on Oct. 22. Its Releases section was empty in that capture, so a Go install from @latest pulls whatever is on the default branch.

What the CLI does

Commands follow a resource pattern, openai <resource> [subresource...] <command> [flags...]. The README example is openai responses create --input "Say this is a test" --model gpt-5.5. Other commands cover file upload and download, image generation and editing, model listing and admin usage reports.

Install is brew install openai/tools/openai or go install 'github.com/openai/openai-cli/cmd/openai@latest'. Standard endpoints read OPENAI_API_KEY, admin endpoints read OPENAI_ADMIN_KEY, and OPENAI_BASE_URL points it at another host. Mutual TLS is listed as beta.

The docs page adds that older Python SDK versions also installed a legacy openai command. A shell that resolves the old binary will not match the guide.

Where the two OpenAI pages differ

The Go requirement is one. Another is image output. The README says openai images generate saves images to ~/Downloads/gpt-images/ by default. The docs page says image commands "do not yet have native --output support," so users must decode the b64_json payload themselves.

Debug logging is a third. The README warns that --debug logs HTTP request and response bodies and that logs may contain sensitive data. The docs say "Authorization is redacted; review headers before sharing logs." Those statements can both be true, since bodies are not headers, but neither page reconciles them.

Topic	README	Docs page
Minimum Go version	1.26	1.25
Image output	Saved to ~/Downloads/gpt-images/	No native --output flag
Debug logs	Bodies logged, may be sensitive	Authorization redacted

Dates in the docs to plan around

The docs page lists two shutdowns. The gpt-4o-mini-tts speech model "will shut down on January 6, 2027", and migration goes through the Realtime API to gpt-realtime-2.1-mini, since changing only the model ID is not supported. whisper-1 and gpt-4o-transcribe-diarize shut down on Feb. 26, 2027.

The README adds two behaviours worth knowing. Interactive runs check GitHub release metadata at most once a day, and the CLI "never installs updates automatically." OPENAI_CLI_NO_UPDATE_CHECK=1 turns the check off.

Pull requests are limited to repository collaborators. Outside users are told to file issues.

How Hacker News received it

The Hacker News thread split. User hahahacorn was "surprised this wasn't available sooner." User bluegatty objected that a "2 trillion dollar company can't put 1 single sentence describing what something is, why it's useful" in the README.

Others asked when to use it instead of Codex, and whether it draws on subscription quota or API credits. The README answers neither question. One commenter, kristopolous, said they built a similar tool called llcat over the past year.

Related coverage includes Google Cloud's private-preview listing of OpenAI models and OpenAI's PHP vulnerability claims.

No release tag exists yet, so the first versioned build is the milestone to watch, along with which Go version the maintainers settle on.