OpenAI Codex Launches Input Prediction Beta for Pro Users
OpenAI Developers announced on the X platform that Codex has launched a beta version of the input prediction (composer predictions) feature for Pro users, which suggests the next message based on the current conversation content and user habits.
Tibo, head of OpenAI Codex, stated that this feature is now available in the desktop application, included in the Pro subscription, and does not consume usage credits. This is the 5th update in Tibo's previously announced "28-day daily iteration" plan.
The feature is currently available to individual ChatGPT Pro accounts (18 years and older), supporting local and SSH threads, with applicable models including GPT-6 Astra and GPT-6.1 Sol. Users can accept suggestions by pressing the Tab key, edit them before sending, and the system will not send automatically.
Predictions are generated solely based on the current thread context and do not independently call memory or other connected applications. Generated predictions themselves do not count against usage credits during the testing period, while sending messages is still calculated according to normal usage. OpenAI claims this feature is one of the most popular new features in internal testing.
Tibo's 28-day plan requires a clear improvement to be launched daily for most Codex or Work users, or a complete reset of credits. Performance optimizations and instant guidance updates have been rolled out in the previous days.
The feature is currently limited to specific threads in the desktop application and has not yet expanded to the web version or all clients. Some users have reported that the suggestion accuracy is high, while others have inquired about frequency control and broader availability.
This feature aims to attract Pro users to continue using it by reducing the input cost of repetitive prompts, benefiting highly active Pro subscribers and OpenAI's user retention, while putting pressure on credit calculations and the pace of feature expansion.
Source: Public Information
ABAB AI Insight
Tibo (Thibault Sottiaux) is the head of the OpenAI Codex product, who has previously maintained user engagement through credit resets and feature updates. On October 4, he announced that over the next 28 days, either a clear improvement valuable to most users would be launched daily, or a complete reset would occur. Input prediction is the 5th improvement in this series.
Resource investment is focused on the composer layer of the desktop application, generating complete next message suggestions using the current thread context rather than word-by-word completion. The feature does not consume credits during the testing period, aiming to reduce the cost of repetitive prompts in the agent loop while locking high-value features at the Pro level.
Similar paths can be seen in other programming agent tools with prompt suggestions or autocomplete features, such as next action predictions in some IDEs. Current AI programming agents are transitioning from pure code generation to reducing human prompt friction and enhancing continuous dialogue efficiency.
Essentially, this is a technological substitution: embedding the prediction of the next human input into the agent loop to reduce repetitive labor. The mechanism allows the agent to master the current context and user habits, pre-filling complete messages after a reply is finished, requiring users only to confirm or fine-tune; limiting this capability to Pro and not counting against credits enhances the experience for high-paying users while avoiding additional pressure on the overall usage pool, and accumulates feedback for broader future availability.
ABAB News · Cognitive Laws
Predicting the next message is less labor-intensive than generating the previous one.
A feature that does not consume credits is a true retention tool.
Delivering daily for 28 days turns improvements into a rhythm.