OpenAI says Codex users can now redirect active tasks instantly

OpenAI product lead Thibault Sottiaux said in an October 8th post on X that Codex users can now redirect an active task "instant" mid-run, letting them course-correct before the model spends more effort in the wrong direction. Sottiaux gave no latency measurement or before-and-after data, and OpenAI's documentation notes that accepted steering input in the API is queued for a continuation and does not undo completed actions or cancel tools already in progress. The post also highlights GPT-6.1 Sol Ultrafast, which OpenAI previewed at its September 29th GPT-6.1 Sol launch as a faster Codex tier capable of up to 300 tokens per second, or up to eight times standard speed, per company-reported peak rates.

OpenAI says Codex users can now redirect active tasks instantly Codex lead Thibault Sottiaux says the change lets users course-correct while a task runs. OpenAI had already previewed GPT-6.1 Sol Ultrafast as a faster Codex tier. By Ryan Merket https://runtimewire.com/author/ryan-merket · Published · Updated Primary source: X https://x.com/thsottiaux/status/2108275041276420573 Why it matters OpenAI is trying to make longer-running agents easier to supervise: faster model output and quicker mid-task corrections may reduce friction, but neither establishes that Codex makes fewer mistakes or costs less on a completed task. Codex users can now redirect an active task more quickly, giving them a way to correct the model before it spends more effort in the wrong direction. OpenAI product lead Thibault Sottiaux described the change in an October 8th post on X https://x.com/thsottiaux/status/2108275041276420573 , calling it "instant." https://x.com/thsottiaux/status/2108275041276420573 https://x.com/thsottiaux/status/2108275041276420573 Mid-task correction is a basic control for agentic coding. Users can discover that a task's scope or approach is wrong after Codex has started acting, leaving them to wait, restart, or let the model continue on a bad path. OpenAI's documentation describes steering as a way to add requirements or change direction while a response is running. In the API, accepted steering input is queued for a continuation; it does not undo completed actions or cancel tools already in progress. Sottiaux did not give a latency measurement or define what "instant" means. OpenAI's public documentation explains how steering works, but his post includes no before-and-after performance data. Developers will need to see whether a correction arrives soon enough to prevent wasted steps in longer jobs. The post also brings attention to GPT-6.1 Sol Ultrafast, which OpenAI had already previewed as a faster Codex tier. At its September 29th introduction of GPT-6.1 Sol, a lower-cost model for coding, computer use and professional tasks, OpenAI said Ultrafast would follow and could generate tokens up to eight times faster than standard speed in Codex. Its DevDay recap later described the tier as coming soon and said it could reach 300 tokens per second. Those are company-reported peak rates, not guarantees that tasks finish eight times faster; tools, reasoning and task length affect elapsed time. OpenAI's initial Sol announcement said the standard model was available in Codex and ChatGPT Work for paid plans and through the API. Sol was not then available in ChatGPT's main chat interface. Sottiaux's October 8th post does not spell out eligible plans or whether the same timing applies across Codex, ChatGPT Work and the API. OpenAI's June report said Codex had more than 5 million weekly active users, with knowledge workers making up about one-fifth of users at that point. As the company pitches Codex beyond software engineers, people supervising longer-running tasks need ways to intervene without restarting work or waiting for a final result. Faster steering could make those corrections less cumbersome, but it does not establish that Codex makes fewer mistakes, follows instructions more faithfully or completes a job at lower total cost. Developers will still need to inspect what Codex changed and verify the result.