Key facts
- GPT-6 Astra is the model that powers dots, and OpenAI also ships it in ChatGPT work, Codex, and the API.
- OpenAI lists API pricing from $10 per million input tokens and $50 per million output tokens.
- OpenAI reports Terminal-Bench 4.0 at 57.9% and DeepSWE v1.1 at 74% — vendor-reported figures.
- Enterprise access to Astra is off by default and enabled by admins; zero data retention is available for eligible API customers on supported endpoints.
dots are the product. GPT-6 Astra is the model doing the work. OpenAI announced Astra the week before DevDay, and the relationship between the two matters: when someone says "dots can do X", the honest expansion is "Astra can do X, and dots give it a persistent job, a computer, and connected apps".
What OpenAI says Astra is
OpenAI describes GPT-6 Astra as its smartest and best-aligned model, available in ChatGPT work, Codex, and the API. Its claimed strengths are specifically the ones an agent needs:
- computer use
- browsing
- professional work
- software engineering
- cybersecurity
- science
The practical claim under those headings is that Astra can not only write code but operate the applications people already use, including applications with no API. That is OpenAI's argument for why an enterprise can deploy it into existing workflows without a data-preparation project first.
The numbers OpenAI publishes
These are vendor-reported results as published by OpenAI on 2026-09-30. They are not independently verified by us, and benchmark conditions change.
| Evaluation | Astra | Comparison OpenAI gives |
|---|---|---|
| Terminal-Bench 4.0 (agentic terminal tasks) | 57.9% | GPT-5.6 Sol 37.3%, Claude Fable 5.1 55.8% |
| DeepSWE v1.1 | 74% | OpenAI describes this as a new record |
| Estimated API cost per task | — | ~9% lower than its own comparison and ~63% lower than the other |
| Business-scenario safety eval (unintended outcomes) | — | 89% less frequent than GPT-5.6 Sol, 74.7% less than Claude Fable 5.1 |
| Financial Modeling World Cup challenge | ~4x faster than the human champion | OpenAI's own comparison |
Read those as marketing with real methodology behind them, not as neutral findings. The safety line in particular is OpenAI measuring its own model against competitors on OpenAI's evaluation, which is useful signal about intent and much weaker signal about relative safety in your environment.
Pricing and access
- API pricing: from $10 per million input tokens and $50 per million output tokens.
- Zero data retention is available to eligible API customers on supported endpoints.
- Enterprise access is off by default and is enabled by an administrator against the applicable price list and agreement.
If you are trying to work out what dots cost from Astra's token price: don't. OpenAI does not bill dots as tokens consumed from your account in the way API pricing implies, and no dot-level credit figures have been published.
The enterprise controls that come with it
Astra's launch shipped governance features that are directly relevant to anyone considering dots in an organisation:
- restrict the model to approved websites and desktop applications
- manage file uploads and downloads
- control browsing history
- confirmation policies that require approval before consequential actions
- automatic review of tool calls that may be unsafe or unauthorised
New enterprise plugins shipped alongside: Oracle Analytics, Power BI (Microsoft Fabric), Navan, and Avalara.
A naming warning
Press coverage in the days around DevDay used several similar model names — variants of "GPT-6.1" and "Astra" — in ways that do not all match OpenAI's own pages. We have deliberately not repeated those claims anywhere on this site. The model OpenAI itself names as powering dots is GPT-6 Astra; where OpenAI's page compares Astra to an earlier in-house model, it calls that model GPT-5.6 Sol.
Outlet headlines from 2026-09-28/29 reference a model release being delayed or abandoned, and a cheaper model that "nearly matches" Astra. We have only seen headlines for those pieces, not the full articles, and the naming is inconsistent. Treat any lineage diagram you see online — including ones that look confident — as unverified until OpenAI publishes it.
Astra versus dots: which one are you buying?
| GPT-6 Astra | dots | |
|---|---|---|
| What it is | A model | Always-on agents built on that model |
| Where you get it | ChatGPT work, Codex, the API | ChatGPT on Pro and Business Premium plans |
| Priced as | Tokens, from $10/$50 per million | Included with the plan; credit terms unpublished |
| Persistent job | No — the caller supplies the context | Yes — the dot keeps a project and a memory |
| Own computer | Only as far as the calling product provides one | Yes, one per dot |
| Typical buyer | Developers and platform teams | Individuals, and teams on Business Premium or Enterprise |
If you are an engineer deciding where to spend effort, that table is the decision: Astra is compute you orchestrate, dots are a worker you supervise.
Safety documentation
OpenAI publishes a deployment safety page for Astra, including static jailbreak evaluations and a page on unintended engagement with external agent messages. We have not yet read it end to end, so we cite nothing from it as fact — it is listed in our source index as pending, and it is the first thing we will read next.
Related reading: what dots are, dots pricing, and the limitations and risks that OpenAI itself documents.
Sources
- OpenAI — GPT-6 Astra: a next-generation model built for work — https://openai.com/index/gpt-6-astra-next-generation-work/
- OpenAI — Introducing dots — https://openai.com/index/introducing-dots/
- OpenAI — deployment safety: GPT-6 Astra — https://deploymentsafety.openai.com/gpt-6-astra
Found an error or a fact that has changed since our last update? Tell us — corrections are published, not quietly edited.