OpenAI Dots: What OpenAI Documented, What Is Reported, and What an Always-On Astra Costs

OpenAI Dots are real, shipped and documented: always-on agents inside ChatGPT, powered by GPT-6 Astra, each with its own cloud computer and browser, able to start and continue Codex tasks, and drawing on context across ChatGPT, Slack and Teams. They were announced at DevDay on 2026-09-29. Two popular descriptions of Dots — that they are Operator v2 under a new name, and that every goal is split across a team of specialist sub-agents — have no OpenAI statement behind them that we could find. What we can add is a measurement: on our executed benchmark GPT-6 Astra scored 9 out of 9 at $8.19 per 1,000 tasks, 68th cheapest of 75 models at that score. An agent that never stops calling that model is billed by what it re-reads on every wake-up, not by what you ask it.

Dots arrived with a cartoon avatar and a lot of secondhand description. Some of that description comes from OpenAI's own documentation, some from reporters who were at DevDay, and some from nowhere in particular. They deserve different amounts of trust, so we sorted them before getting to the part we can measure.

The ledger: confirmed, attributed, unsourced

Confirmed means OpenAI's own documentation says it — here, the Meet dots page in OpenAI's ChatGPT Learn docs, read 2026-10-03. OpenAI's announcement post, Introducing dots, refused our fetch, so where we cite its wording we cite it through an outlet. Attributed means a named outlet reports it, read on 2026-10-03. Unsourced means it is circulating and we could not find anyone's name on it.

The Operator story is understandable: OpenAI's previous browser agents were retired, and Dots does browse. But a product filling a gap is not the same product renamed. Wikipedia's Operator article, read 2026-10-03, does not list Dots as a successor, and OpenAI's Dots documentation does not mention Operator at all. The specialist framing is closer to something real: OpenAI documents parallel background agents, and TechCrunch reports that OpenAI envisions specialist dots. Neither says a coordinator splits every goal among specialists.

What coordinating Codex actually means

The documented version is narrower than the hype and more useful. Per OpenAI's page, a dot can start Codex tasks on a computer you have connected and pick up local Codex tasks already in progress, and those tasks show up on that machine. The dot is the dispatcher; Codex does the coding. That split matters for cost too: the Codex tasks a dot starts count toward Codex's usage limits, not the dot's. A dot that hands out a lot of coding work spends two budgets.

TechCrunch's reporter notes the functionality overlaps with what Codex already does, with Dots bundling it around independent action. That is a fair read of the documentation.

The model underneath, measured

We cannot benchmark a dot; it is a ChatGPT product, not an API endpoint. We can benchmark the model OpenAI says runs it. GPT-6 Astra went through our nine executed Python tasks on 2026-09-16, priced at the list rate captured 2026-09-15. Its GPT-6 siblings ran on 2026-10-02.

Astra itself is well behaved here: 47 reasoning tokens per call, 5.6 seconds, and it billed exactly the 604 input tokens we sent, same as every sibling. It is also about 51 times the cost of GPT-6 Luna ($8.19 ÷ $0.16) for the same nine passes, and 273 times Solar Mini 4, which as of 2026-10-03 is the cheapest and fastest model to score 9 out of 9 on our board, at $0.03 and 2 seconds.

Be clear about what that does and does not mean. Nine self-contained Python functions cannot separate a frontier model from a competent small one — 75 models clear this suite. OpenAI did not put Astra inside Dots to write two_sum; it put it there for long, multi-step, judgement-heavy work our suite never exercises. What the suite does give is a dated baseline: an observed cost on these particular prompts, which is $0.00819 per task on our prompts ($8.19 ÷ 1,000).

Why overhead becomes the bill

A chat model is billed for what you send. An always-on agent is billed for what it sends, and most of that is not your request. Per OpenAI's own description, each time a dot does something it can draw on the conversation, relevant ChatGPT memory and its own saved notes, across several channels, with a plugin surface reported at over 4,000 apps. It also decides when to pause and wake, and can run background agents in parallel. Every one of those wake-ups and sub-runs is a model call that starts by re-reading standing context.

We have watched this pattern on the API already. On our bench, plain GPT-6 models billed only what we wrote. Pro-tier endpoints did not: the one Pro entry on today's data sheet, GPT-5.6 Luna Pro, read 21,729 input tokens on the same nine prompts every GPT-6 model answered with 604. We measured the same thing on GPT-6 Astra Pro and the other GPT-6 Pro endpoints — the full numbers are in our GPT-6 Pro write-up and the GPT-6 Astra review. The mechanism is provider-side scaffolding prepended to every call, which we explain in hidden system prompt tokens. On short requests the scaffolding, not the question, is most of what you pay for.

Dots is that pattern turned into a product: scaffolding is the point. Here is the unit conversion that makes it concrete, on Astra's list price of $10 per million input tokens (2026-09-15): every 1,000 tokens of standing context a dot re-reads costs $0.01 per wake-up. That single re-read is already more than one entire benchmark task on Astra, at $0.00819. A wake-up that checks an inbox, finds nothing and goes back to sleep still pays for the read. Caching cuts the rate on repeated context but does not make it free — see prompt caching. And output is the expensive side: Astra lists output at five times input ($50 ÷ $10), so a dot that writes long status reports nobody reads pays the higher rate.

We modelled how turns, context and reasoning compound in a running agent in what an AI agent actually costs. The short version applies here: in an agent, per-call overhead multiplied by call count is the bill.

What a dot costs you

For a ChatGPT subscriber, not per-token money — at least not yet. OpenAI's documentation says conversations with a dot do not count toward ChatGPT usage limits, that plans include an allowance for deeper work, and that limits are extended for the first month after launch. 9to5Google reports one dot is included with Pro and Business Premium and more will be purchasable later; DataCamp reports OpenAI will share per-plan usage terms after the first month. No per-dot price had been published as of 2026-10-03, and we will not guess one.

That makes the overhead OpenAI's cost for now and your allowance later. If you are building a dot-like agent yourself on the Astra API, it is your cost from the first call: the numbers above are one measured baseline, and the standing context you design is the multiplier. Whether the work needs Astra at all is worth testing — on our suite, GPT-6.1 Sol and GPT-6 Luna passed the same nine tasks for a fraction of the price, although our suite is not the work Dots is sold for.

How these numbers were produced

Nine Python tasks, each a function signature plus a spec and no example tests. Generated code executes against hidden asserts in an isolated python3 -I subprocess with a 12-second timeout. Temperature 0, max_tokens 4000, one scored attempt per task. An API-layer failure is recorded separately from a wrong answer. Cost is derived — measured input and output tokens multiplied by the list price captured on the date shown — not a billing statement. GPT-5.6 Luna Pro was priced at $0.5 in and $3 out on 2026-08-06 and its list price has moved since. Runs go through OpenRouter, not the DataLLM Lab gateway. Third-party facts above were read on 2026-10-03. Full method on the methodology page, and prices move.

What we did not measure

• A dot. Dots are not exposed as an API endpoint, so nothing here is a measurement of the product itself.
• How much context a dot re-reads per wake-up. The $0.01 per 1,000 tokens figure is arithmetic on list price, not an observation of Dots.
• Long-horizon agent quality. Our suite is nine single-turn functions; it cannot rank models on the work Dots is built for.
• How OpenAI meters dots internally. Subscription allowances are not token bills, and OpenAI has not published the post-launch terms.
• Repeat runs. One scored attempt per task; single-run figures, not averages.

Evidence: dated measurements and catalogue fields; method, latency statistics and limitations. Measurements describe these runs, not all providers or future versions.

Billing scope: our Astra API experiment does not measure Dots subscription allowances, its internal prompts or actual Dots operating cost. Overhead discussion is an inference for API-built agents, not an observed bill for this product.

Primary sources, checked October 3, 2026: OpenAI: Meet dots; OpenAI release notes. Dated measurements above may differ from the current documentation.