On August 27, 2026, WIRED reported that OpenAI is testing a new “Persistent mode” for Codex, its command-line coding agent. The report, by Maxwell Zeff, is based on a review of code in OpenAI’s public Codex repository rather than a product announcement. As described, Persistent mode is a new setting in Codex’s reasoning-effort menu that lets the agent keep working on a task until a user explicitly puts it to sleep, instead of stopping at the minutes-to-hours ceiling current runs hit. OpenAI has confirmed it is exploring the feature but says there are no immediate plans to launch it.
If the term turned up in a headline and the question is simply what it is, whether it is real, and whether it can be used today, the short answer is that it is real as code and as a stated area of exploration, but it is not a released feature. The rest of this piece keeps four things separate that coverage tends to blur: what OpenAI has confirmed, what WIRED reported from the code, what can be checked directly in the public repository, and what is still unknown.
What Persistent mode reportedly is
Codex already exposes a reasoning-effort setting, which controls how much reasoning the model applies to a task, and Codex offers multiple reasoning-effort levels. According to WIRED, Persistent mode appears as a new option in that same menu, but rather than dialing reasoning up or down it changes the stopping condition: the agent is told to continue working until it is put to sleep. As reported, the distinction is between increasing reasoning effort for a bounded task and a mode intended to continue until the user puts it to sleep. The reported framing is a mode that can spend more compute, tokens, and time than the current settings precisely because it is not designed to halt on its own.
Where the information came from
The origin for the specific claims is WIRED’s August 27 report, grounded in code the outlet reviewed in OpenAI’s Codex repository — the command-line version of the agent. Other outlets covering Persistent mode so far are restating that report. The distinction matters for calibration: the prompt-level details below — proactivity, “knowledge of the user,” unprompted messaging — come from WIRED’s reading of the code, not from OpenAI product documentation, and OpenAI has not published a specification for any of them. Two parts of the story, though, can be checked without relying on the reporting at all.
What can be verified in the public Codex repository
The openai/codex repository is public, and the plumbing the reporting points to is visible in it. A pull request titled “Support persistent reasoning effort” (#40799) was merged on August 26, 2026; it adds a persistent value to the reasoning-effort protocol and the TypeScript SDK types, wires it through parsing, serialization, and the terminal UI’s reasoning selector, and maps the new setting onto the Responses API. Separately, an earlier change, “Add interruptible sleep tool” (#28429), merged in June 2026, gives the agent a built-in sleep tool that pauses a turn and resumes when new input arrives. Neither pull request proves the full behavior WIRED describes, but both confirm that the scaffolding for a persistent, sleep-and-wake agent is genuinely in the codebase rather than inferred.
What Proactivity is, and how it differs from Persistent mode
Persistent mode governs how long a run lasts; Proactivity, per WIRED, governs what the agent chooses to do. The reviewed code reportedly describes a separate feature that instructs the agent, after finishing a request, to create its own follow-up tasks and work on them across sessions — drawing on past interactions and “knowledge of the user” to decide what to pick up next. The same code is said to describe a tool that lets the agent message a user without being asked, with an instruction to do so sparingly. Persistence keeps the agent running; proactivity determines what it may choose to work on next. The reported code also includes a guardrail: Persistent mode does not expand what the agent is permitted to do, and changing anything outside the user’s own system still requires approval first.
What OpenAI has actually confirmed
OpenAI’s on-record position is narrow. A spokesperson told WIRED the feature is being explored in the shared Codex codebase — which the company noted could let it reach products beyond the command line — and that there are no immediate plans to launch it. OpenAI also acknowledged that persistent AI systems carry heightened risks. That is the extent of the official confirmation: exploration, a shared codebase, no launch timeline, and a general note on risk. It is not a product announcement, a beta, or a dated roadmap, and OpenAI has not adopted any “always-on agent” branding — that phrasing comes from headlines, not the company.
How this differs from Codex today
Long-running Codex is not itself new. OpenAI has publicly demonstrated Codex models working autonomously for extended stretches — a widely cited example ran for roughly 24 hours on a single task, using on the order of 13 million tokens and generating tens of thousands of lines of code, made possible by “compaction” that carries a task across many context windows. But those runs are bounded: they pursue one assigned task and stop when it is done or the budget is exhausted. The reported novelty of Persistent mode is twofold — removing the automatic stop so the run continues until a user ends it, and, through proactivity, letting the agent choose its own next tasks rather than only executing the one it was given. That is closer to an agent that stays on than to a scheduled or cron-triggered job, because the agent, not a timer, decides the work.
The open questions
The gap between a bounded run and one with no automatic end raises questions the code does not answer. An agent that keeps working, and generates its own follow-up work, spends compute and tokens for as long as it stays awake — so how such runs would be metered, capped, or billed is an obvious implication to weigh, but it is a question, not a disclosed OpenAI behavior. Nothing public states how Persistent mode would treat token usage, what limits would apply, which plans or models would get it, or how the sleep-and-wake control would be surfaced to users. Those are the details that usually decide whether a feature like this is usable in practice, and none of them exist yet. The table below collapses the same picture into one reference.
| Detail | What we know | Status |
|---|---|---|
| Persistent reasoning-effort option in the code | Merged into openai/codex on August 26, 2026 (PR #40799) | Verifiable in repo |
| Interruptible sleep tool | Merged into openai/codex in June 2026 (PR #28429) | Verifiable in repo |
| OpenAI is exploring the feature | Confirmed by an OpenAI spokesperson to WIRED | Confirmed |
| “Continue working until put to sleep” | Described in code reviewed by WIRED | Reported |
| Proactivity — self-created follow-up tasks across sessions | Described in code reviewed by WIRED | Reported |
| Unprompted user messaging (used sparingly) | Described in code reviewed by WIRED | Reported |
| No expanded permissions; external changes need approval | Instruction in the code reviewed by WIRED | Reported |
| Release date, pricing, token billing, limits, eligible plans, final behavior | Not stated anywhere public | Unknown |
Is Codex Persistent mode available now?
No. As of August 28, 2026, Persistent mode is not a released Codex feature. It exists as code in OpenAI’s public Codex repository and as a feature OpenAI says it is exploring, but the company has stated there are no immediate plans to launch it. There is no public beta, no announced availability by plan, and no release date.
What is Codex Proactivity?
Proactivity is a separate feature described in the code WIRED reviewed. It reportedly instructs the agent to create its own follow-up tasks after completing a request and to work on them across sessions, using past interactions and “knowledge of the user” to choose what to do next, and it includes a tool for messaging the user unprompted, to be used sparingly. It is distinct from Persistent mode: persistence governs how long a run lasts, while proactivity governs whether the agent generates its own work. Both come from WIRED’s code review, not from official OpenAI documentation.
How is Persistent mode different from Codex’s existing long-running tasks?
Codex already runs long autonomous tasks — OpenAI has shown runs lasting around a day, enabled by compaction across context windows — but those runs target one assigned task and stop when it finishes or the budget runs out. Persistent mode, as reported, changes the stopping condition so the run continues until a user puts it to sleep, and pairs with proactivity so the agent can choose its own next tasks. The difference is an agent that stays on and self-directs, versus one that completes a defined job and halts.
Did OpenAI officially announce Persistent mode?
No. OpenAI confirmed to WIRED only that it is exploring the feature in its shared Codex codebase and that there are no immediate plans to launch it, along with a general acknowledgment that persistent systems carry heightened risk. It has not published product documentation, branding, or a roadmap for Persistent mode. Descriptions of its behavior come from code review and reporting, not an official announcement.
Agentic coding tools keep multiplying, and each one adds usage-based AI spend on top of the cloud bill — spread across providers and easy to lose track of. CloudQuell puts your AWS bill next to itemized OpenAI and Anthropic spend on a single flat-priced ledger, so the cost of the tools your team runs stays visible in one place. Free under $10K/month of tracked spend.
Try CloudQuell →