OpenAI's Dots run free until they open a Codex task

"I got community noted, but the note is wrong," Tibo Sottiaux wrote on X. He was defending OpenAI's claim that its new always-on agents, called Dots, can run around the clock without eating into your subscription. As The New Stack reports, that holds only until a Dot hands work to Codex, and only for as long as OpenAI's launch terms last.

At a glance

• OpenAI's head of core products and platform says each user's primary Dot is included in the plan, available 24/7, and spends none of the plan's allowance when it does work directly.
• Work a Dot starts in Codex or ChatGPT Work counts against those products' limits, and the $200 Pro plan's included usage was halved on launch day, alongside a new $500 tier.
• Published terms keep Dot usage off eligible plan allowances only for the first month after launch, and Sottiaux did not say whether a Dot warns you before moving work into metered products.

If you missed the earlier episode: according to Yahoo Tech, Sottiaux once admitted that Codex was doing more behind the scenes than intended, with helper subagents running twice or retrying too aggressively, and OpenAI reset limits for all users. The same outlet notes that AI providers have been moving away from all-you-can-eat use of their tools, and that Anthropic adjusted Claude usage caps during peak hours in March.

Sottiaux says the baseline Dot will always be included, but the published terms cover one month

Sottiaux's launch post drew a Community Note asking whether always-on agents could realistically run inside existing subscriptions without blowing past usage limits. His reply laid out the rule. The primary Dot is part of the plan, works 24/7 and draws nothing from the allowance while it does the work itself. He went further and said the baseline "will always just be included in your plan."

OpenAI's published terms are narrower. They say Dots usage won't count toward eligible plan allowances for the first month after launch, and that per-plan usage terms will be published afterward. Developers sizing workloads around Dots therefore don't yet know what the included allowance will look like once that window closes.

The meter starts the first time a Dot opens a Codex task

The exception is delegation. If you ask a Dot to create a task in Codex, that task draws usage as usual, and the same goes for work a Dot starts in ChatGPT Work. A Dot can sit online all day without touching your allowance, then begin spending it the moment it opens a Codex task.

The allowance itself is shrinking. On the day Dots launched, OpenAI halved the included usage on its $200 Pro plan and added a $500 tier above it. Dots went to the Pro tier first; TechCrunch adds Business Premium users in eligible markets, and WIRED says users start with one Dot, with multiple agents expected later.

Scale costs OpenAI money too. Sottiaux said the company expects "a few million dots online within days," and running that many agents continuously costs money even when they never call another product. For the launch period, OpenAI is absorbing that baseline activity into the subscription for each user's primary Dot.

A Dot chooses where work runs, and that choice sets the bill

According to WIRED, Dots run on GPT-6 Astra and keep crawling the web, pulling context from connected apps and learning your preferences over time. TechCrunch says much of this was already possible through Codex and similar agentic harnesses, and that Dots bundles it into a package built for independent action, which is why Dot work overlaps with Codex.

A Dot can often finish the same job several ways: do it itself, call an included tool, or send it to a metered product like Codex. It makes that call without anyone reviewing each step, so the route can decide whether you are charged at all. Picture a contractor on a flat retainer who can fix the leak himself or call in a plumber billed to your card.

OpenAI's Codex pricing page on ChatGPT says local messages and cloud chats share the plan's usage allowance and that weekly limits may also apply. Codex is also not the only place delegated coding work can land: AWS is pitching an open source agent it says costs 45% less to run than Claude Code or Codex.

Sottiaux didn't say whether users will be warned before a Dot moves from included work into metered usage, so you may only find out when you check what's left. The approval prompts WIRED describes cover sensitive actions such as installing software or changing a password, and nothing in the reporting ties them to spending. Oddly, the one decision that moves money appears to have no stated checkpoint, and Yahoo Tech's account of Sottiaux's earlier Codex episode shows how hard background usage is for users to see.

When the one-month window closes

OpenAI says it will publish per-plan usage terms after the first month, but has given neither a date nor any numbers. Sottiaux has also previewed a paid layer: "In the future you will be able to increase the speed and allow your dot to have more bandwidth." The extra capacity will cost money while the baseline stays in the plan. No price for that layer has been named.