OpenAI Dots run on GPT-6 Astra. That is the model OpenAI names on the Dots launch post, the Dots safety post, and ChatGPT’s Meet dots page. Astra is also the model OpenAI already shipped in ChatGPT Work, Codex, and the API.

This page stays on what OpenAI published about Astra, then ties those facts to Dots. It does not treat benchmark tables as a promise that your Dot will hit the same score on your laptop.

#What is GPT-6 Astra?

OpenAI’s model page calls GPT-6 Astra “the world’s most intelligent and aligned model.” It says Astra is state of the art on computer use, browsing, software engineering, cybersecurity, science, and professional work.

The work-focused Astra page (fetched September 29, 2026) says the model is available in ChatGPT Work, Codex, and the API, and that in those products it can write code and work through the same applications people use even when those applications have no API.

On the same day as Dots, OpenAI’s DevDay recap introduced GPT-6.1 Sol as a cheaper model with “near-Astra intelligence” at a fifth of Astra’s standard input and output token prices. The recap lists Sol for API, Plus, Pro, Business, Enterprise, and Edu. Dots documentation fetched that day still says Dots are powered by GPT-6 Astra.

#Why does Astra matter for Dots?

A Dot is supposed to keep working on a cloud computer, use a browser, call plugins, and hand work to Codex or ChatGPT Work. That is computer use plus long-running tool use, which is the job OpenAI trained Astra for.

The Dots safety post says Astra is trained to understand your goal, stay inside the scope of the request, and ask focused questions when the answer could change what it should do. It says OpenAI trained Astra to refuse harmful requests, including biological or cybersecurity misuse, and tests Dots on changing instructions, ambiguous requests, and attempts to push past permissions.

ChatGPT Learn describes the product result: the Dot lives in the cloud, has its own computer and browser, and can research, analyze data, prepare documents, and build software while you are away.

#What scores did OpenAI publish?

All figures below are from OpenAI’s GPT-6 Astra page, retrieved September 29, 2026. The page says evaluation scores are the maximum at any effort, and that research or API runs can differ from production ChatGPT.

Computer use and browsing:

Evaluation (OpenAI’s table name)

GPT-6 Astra

GPT-5.6 Sol

OSWorld 2.0 (v2026.08.08, offline set, partial score)

72.6%

65.7%

Agents' Last Exam

59.3%

53.6%

ScreenSpot-Pro (no tools)

92.7%

76.9%

BrowseComp

91.5%

90.4%

OpenAI also writes that in latency simulations on OSWorld 2.0, Astra scored 72.6% at roughly 40 minutes per task, compared with 65.7% at roughly 75 minutes for GPT-5.6 Sol, about 47% less time per task.

Coding and professional work:

Evaluation

GPT-6 Astra

GPT-5.6 Sol

Terminal-Bench 4.0

57.9%

37.3%

DeepSWE v1.1

74.1%

72.7%

AutomationBench

41.4%

18.1%

BenchCAD

95.9%

83.3%

The intro of the same page says Astra “saturates FrontierMath Tier 4 with a 98% score.” The evaluation table on that page lists FrontierMath Tier 4 (v2) at 97.6%. This article uses the table’s 97.6% for the named eval. ARC-AGI-3 is 99.9% in both the intro and the table. ExploitBench is 100.0% in the table.

Alignment and safety (lower is better unless noted):

Evaluation

GPT-6 Astra

GPT-5.6 Sol

Internal computer use safety benchmark

2.4%

22.0%

Same benchmark with AutoReview

1.8%

4.3%

Internal circumvention benchmark

0.00%

0.29%

ExploitGym honeypot

0.0%

48.2%

Internal hallucination benchmark

4.2%

12.2%

The work page says that on an internal computer-use safety benchmark, Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1. The intro describes a Hugging Face-inspired eval: GPT-5.6 Sol without production safeguards went beyond the authorized target 48% of the time; GPT-6 Astra did this in 0% of cases.

OpenAI says Astra never attempted to circumvent a Codex Auto-review denial in an internal evaluation, even when Auto-review was configured to be evadable.

#How much does Astra cost?

On the GPT-6 Astra page, OpenAI API Standard pricing is $10 per million input tokens and $50 per million output tokens. Separate rates apply to cache reads and writes. Fast mode is available in the API and is described as up to 2 times the speed of Standard at 2 times the Standard price. The model id is gpt-6-astra.

That is API list pricing. Dots are a ChatGPT feature. OpenAI says conversations with your Dot do not count toward ChatGPT usage limits, while Work and Codex tasks the Dot starts still count as usual. Do not treat the $10 / $50 API numbers as the price of a Dot.

DevDay’s recap adds Ultrafast: up to 8 times faster token generation (300 tokens per second) in Codex and up to 6 times in the API. GPT-6 Astra Ultrafast is listed as available in the API and in ChatGPT Work and Codex on Pro 500 and Enterprise plans. GPT-6.1 Sol Ultrafast is listed as coming soon. Dots pages do not say the Dot itself is billed as Ultrafast.

#How does Astra’s cybersecurity rating affect Dots?

The Astra page says Astra is the first model to reach the Critical cybersecurity capability threshold under OpenAI’s Preparedness Framework. Without production safeguards, OpenAI reports 100% on ExploitBench and 42.4% on ExploitGym, and says expert-led tests found Astra could use previously unknown vulnerabilities to get arbitrary code execution in hardened browsers.

The shipping product refuses more advanced cybersecurity tasks such as creating proof-of-concept exploits. OpenAI says extra safety checks can slow, pause, or stop legitimate work, including defensive cybersecurity. In ChatGPT or Codex you may be asked to review an action; in the API the task stops.

Dots add another layer on top of that model: a sandboxed cloud computer, Auto-review, Custom Rules, and monitoring that can pause or stop the Dot. OpenAI’s Dots safety post still says Dots can make mistakes. Review consequential work.

#Where else does Astra run?

Availability on the Astra page (retrieved September 29, 2026): rolling out to ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API, Microsoft Azure, and AWS Bedrock. Usage is included in existing subscription allowances, with credits for extra use. Pro, Business, and Enterprise also get GPT-6 Astra Pro. Enterprise administrators enable Astra; access is off by default at that model’s launch.

Zero Data Retention is available for eligible API customers on supported endpoints, subject to approval. That API control is separate from Dots. OpenAI’s local-computer-access guide says Dots in the Enterprise beta do not provide strict zero data retention.

#What to watch next

Watch whether OpenAI keeps Dots on Astra or offers Sol or Ultrafast as a Dot speed setting. The launch post already says you will later scale a Dot by increasing its speed or its monthly work. As of September 29, 2026, the documented model behind Dots is still GPT-6 Astra. For the product wrapper, see what OpenAI Dots are. For the rest of DevDay’s model news, see the DevDay 2026 announcements.