Codex Desktop Adds Free Composer Predictions in Pro Beta

A new beta feature suggests next prompts via Tab in local GPT-6 Astra and Sol threads, with no extra credit charges for ChatGPT Pro users.

OpenAI rolled out a beta update on October 9, 2026, introducing contextual prompt autocomplete to its Codex desktop client. Known as composer predictions in Codex Desktop, the system generates inline text recommendations inside the prompt composer immediately after the assistant finishes an answer. The test phase is currently limited to individual ChatGPT Pro subscribers aged 18 and older, as outlined in the OpenAI Help Center and official OpenAI Release Notes.

The feature name initially caused confusion with the Composer multi-file editing mode in Anysphere’s Cursor editor. The update bears no connection to Cursor. In OpenAI’s ecosystem, “composer” refers strictly to the text input component where users draft messages inside the Codex desktop interface.

How Composer Predictions Work in Codex Desktop

The predictive engine reads the active conversation thread and configured local memory to anticipate the developer’s next instruction. When an engineer reviews an assistant’s code patch, the interface might suggest running the test suite or applying a syntax fix.

Suggestions appear as inline gray ghost text. Tapping the Tab key accepts the proposal and fills the input bar, but the message does not send on its own. Developers can edit the draft, delete unwanted parameters, or ignore the ghost text entirely by typing over it.

Users who prefer writing prompts from scratch can disable the autocomplete mechanism under Settings > General > Composer > “Show predictions”. OpenAI notes that the system only references open session context and local memory flags, without performing background scans of files on the host computer.

Supported Models and Usage Caps

Composer predictions operate exclusively within local directory threads or remote sessions established over SSH, paired with GPT-6 Astra or GPT-6.1 Sol models. Standard ChatGPT web interfaces and mobile applications remain excluded from the beta.

Throughout the testing phase, generating and reviewing predictions incurs no additional fees and does not count against Codex rate limits. Usage quotas from the Pro plan apply only after the developer accepts the prediction and presses Enter to dispatch the prompt.

Future pricing and availability remain unconfirmed. OpenAI has not stated whether predictions will stay free once the beta period concludes. Similarly, any eventual rollout to Plus, Team, Enterprise, or Free tiers remains a RUMOR, as current documentation explicitly confines access to individual Pro accounts.

Reducing Friction in Agent Workflows

In agent-driven development, developers spend substantial time typing short follow-up prompts to guide execution cycles. Writing repeated requests to inspect logs, execute test scripts, or commit changes introduces constant pauses. Bringing inline completion up to the prompt level helps narrow that gap.

With competition mounting across tools like GitHub Copilot, Cursor, and Claude Code, OpenAI is aiming directly at workflow momentum. Inserting the next prompt with a single keystroke keeps developers focused on reviewing output rather than composing routine commands.