OpenAI Codex Composer Predictions Suggest Your Next Message for ChatGPT Pro on GPT-6 Astra and GPT-6.1 Sol, Free During Beta

OpenAI put composer predictions into beta in the Codex desktop app on October 9, 2026. After Codex finishes a reply, the message box can show a suggested follow-up based on that thread. Press Tab to accept it, then review or edit before you send.

The company’s product release notes and Help Center page say the beta is limited to personal ChatGPT Pro users aged 18 and older in all supported regions. It works in local and SSH threads when you select GPT-6 Astra or GPT-6.1 Sol. Generating a prediction does not count toward Codex usage limits or consume credits during the beta. Messages you actually send, including accepted predictions, follow normal Codex usage and billing.

That is a small quality-of-life feature on paper. For anyone who already lives in Codex all day, it is also a bet that the next prompt is often as predictable as the next autocomplete token, and that saving those seconds compounds.

What OpenAI confirmed

OpenAI’s documentation is unusually explicit about who gets the feature and what it will not do.

The Help Center also draws a line around context. Predictions use the conversation already in the current thread. They do not independently look up memories or pull data from connected apps. If memory context is already present in the thread, it can shape a suggestion, and you can still use predictions with memory off.

OpenAI is clear that a prediction will not appear after every response. Waiting several seconds is normal, and long threads can take longer. You can keep typing without waiting. During the beta, suggestions are full next messages, not word-by-word completion as you type.

How the workflow actually feels

The intended loop is simple. Update the Codex desktop app, sign in with a personal Pro account, open a local or SSH thread on Astra or GPT-6.1 Sol, and send a normal request. When Codex finishes, watch the composer. If a suggestion appears, Tab accepts it into the box. You still decide whether to send, edit, or ignore it.

That last point matters more than the marketing line. Composer predictions do not auto-continue a task. They do not accept or send follow-ups on their own. OpenAI’s FAQ says you should review each prediction before choosing to send it. In other words, this is autocomplete for your next instruction, not an unsupervised agent loop.

That design choice sits next to another Oct 8 release-notes item: faster steering in Codex, so a follow-up can redirect a running task sooner. Predictions help you draft the next steer. Steering is what happens after you send it. Together they tighten the human-in-the-loop path without removing the human.

For builders comparing coding agents, this also sits beside the safety story we covered when researchers tested whether coding agents would erase their own logs. A feature that forces a Tab review before send is a small control surface. It will not fix agent misbehavior, but it does refuse to pretend that every suggested follow-up should fire automatically.

Price, limits, and what “free” means

OpenAI’s wording is careful. During the beta, generating predictions is included at no additional cost for eligible Pro users and does not consume Codex limits or credits. The moment you send a message, normal Codex usage and billing apply, whether you wrote the text or accepted a prediction.

That distinction is easy to miss in social posts that stop at “free.” The free part is the suggestion generation. The billable part is still every message that leaves the composer. If you Tab-accept a long follow-up and send it, you pay for that turn like any other Codex turn.

OpenAI has not published a separate token price for prediction generation itself, because the beta frames generation as bundled and non-metered for eligible Pro accounts. It also has not said whether that free generation survives after beta. Treat the zero-cost generation claim as beta-scoped until the Help Center changes.

For context on the wider GPT-6 stack these models sit in, see our notes on GPT-6 reaching free ChatGPT users with Intelligent UI and the earlier DevDay 2026 roundup on Dots, Sol pricing, and Ultrafast. Composer predictions are a desktop Codex feature, not a ChatGPT chat-tab feature.

Where it fits against Decisions and other OpenAI surfaces

OpenAI is shipping two different “decision” ideas in the same week. The Decisions API is a typed API that returns predicates, choices, and scores for your code. Composer predictions are a UI hint for your next natural-language instruction inside Codex.

One is for programs. One is for people who talk to an agent. Confusing them will lead to the wrong expectation. Predictions will not return calibrated probabilities. Decisions will not Tab-fill your Codex composer.

OpenAI also separates composer predictions from the suggestions that appear on a new conversation page. Those help you start a task. Composer predictions help you continue one after Codex has already answered.

What it means for Indian developers

If you already pay for personal ChatGPT Pro and use the Codex desktop app over local projects or SSH into a VPS, this is an update you can try today without a new vendor. India sits inside OpenAI’s supported-region framing for the beta, subject to the same age and plan rules as other regions.

Practical upside is speed on repetitive agent loops: ask for a failing test, accept a suggested “run it and patch,” then edit the patch scope before send. Practical risk is the same as any autocomplete. A confident wrong suggestion can waste a billed turn if you Tab and send without reading.

Bandwidth and latency still matter on Indian broadband and mobile hotspots. OpenAI says predictions can take several seconds and longer in long threads. You do not need to wait before typing, which is the right default when the suggestion is late.

Teams on Business or Enterprise should not assume they are covered. The Help Center says the beta is for personal ChatGPT Pro in the Codex desktop app. Work may show predictions only in local Work threads. If your company standardizes on workspace seats, wait for an explicit admin-facing note rather than extrapolating from Pro personal.

What remains open

Three caveats stay live. First, free generation is a beta promise, not a permanent price card. Second, availability is narrow: personal Pro, desktop Codex, Astra or GPT-6.1 Sol, local or SSH. Third, OpenAI has not published quality metrics for how often predictions match developer intent, only that they can miss and should be reviewed.

Even with those gaps, the product direction is readable. Codex is collecting small controls that keep a human in the loop while cutting idle time between turns: faster steering on Oct 8, composer predictions on Oct 9. The companies that win coding-agent mindshare may not be the ones that remove the human first. They may be the ones that make the next human message cheaper to write.

Frequently Asked Questions

What are Codex composer predictions?

They are a beta Codex desktop feature that can suggest your next message after Codex finishes responding, based on the current thread. Tab accepts the text into the composer without sending it.

Who can use composer predictions today?

OpenAI says personal ChatGPT Pro users aged 18 and older in supported regions, on the latest Codex desktop app, in local or SSH threads with GPT-6 Astra or GPT-6.1 Sol.

Do composer predictions cost money?

During the beta, generating predictions does not add a separate charge and does not consume Codex usage limits or credits. Sending a message, including an accepted prediction, follows normal Codex usage and billing.

Can predictions run my next step automatically?

No. OpenAI says predictions do not automatically accept or send follow-ups. You review, edit if needed, and send yourself.

How do I turn them off?

In the Codex desktop app, open Settings, go to General, then Composer, and turn off Show predictions. You can turn the same switch back on later.