GPT-6 Astra: what it can do and what people are building with it

GPT-6 Astra is OpenAI's model for demanding coding, research and tool-assisted work. Since its launch, software, browsing and computer-use demonstrations have raised a more useful question than whether it is “the most powerful AI”: what can it actually do, where can people try it, and which outside projects make its abilities verifiable?

What GPT-6 Astra is

GPT-6 Astra is an AI model from OpenAI. ChatGPT is the product that offers different experiences and models. Codex is the software-focused environment; ChatGPT Work is designed for longer assignments and deliverables; regular Chat is for conversation. An experience using Astra does not make those names interchangeable or guarantee that every feature is available in every interface. For a broader comparison of the products, see our ChatGPT vs Claude vs Gemini comparison.

OpenAI announced Astra on 3 September 2026, first rolling it out to a limited set of organizations and then to eligible accounts. The developer model reference gives its API name as `gpt-6-astra`; it documents text and image input and tools such as web search, code interpreter, shell and computer use. Those tools must be enabled by a product or application: the model does not automatically gain access to your files, browser or desktop.

Where GPT-6 Astra is available

As of this cut-off, Astra is available through three main routes. In ChatGPT Work and Codex, eligible plans and permissions may show Astra in the model picker; OpenAI says Plus includes access to those experiences. In Chat, the option called GPT-6 Pro is powered by Astra and is listed for Pro $100 and $200, Business and Enterprise; Enterprise access also depends on admin settings. It is not the default model for every conversation: Chat has its own model options and limits. The API exposes `gpt-6-astra` with billing and rate limits separate from a ChatGPT subscription. Check the current ChatGPT model availability and Work and Codex pages before subscribing or integrating the model. OpenAI notes that Work is still rolling out gradually and Plus has a limited allowance; Pro $500 is not listed among plans eligible for GPT-6 Pro in Chat.

OpenAI describes Work as an agent experience for multi-step assignments and finished deliverables; Codex remains focused on software and technical work. In both, what Astra can do depends on the files, permissions, tools and limits provided. API organizations must configure tools and stay within their usage limits. Azure and AWS Bedrock were named in the launch announcement, but that does not mean every country, plan or external deployment has identical terms.

Astra, GPT-6 Pro, Sol and Luna are different names

OpenAI's naming mixes models and product options. GPT-6 Astra is the model. GPT-6 Pro is a ChatGPT option that documentation says is powered by Astra, but it has its own access path and limits. GPT-6 Sol and GPT-6 Luna are other models in the family announced on 22 September; GPT-6.1 Sol arrived on 29 September as a separate variant aimed at complex work with a different capability-and-cost balance. GPT-5.6 Sol is an earlier generation and remains selectable in some products. These labels are not interchangeable: a Codex result with Astra does not prove how regular Chat will answer with Sol.

Compared with GPT-5.6 Sol, Astra's launch page reports improvements on several tests, but not a universal measure of “intelligence”. OpenAI published 57.9% versus 37.3% on Terminal-Bench 4.0, which evaluates agents on complex terminal tasks such as software engineering, configuration and data analysis. On OSWorld 2.0, an offline computer-use test with a partial score, it reported 72.6% for Astra and 65.7% for GPT-5.6 Sol. These are vendor results on specific tests, not success rates for every real project. OpenAI's table says it uses the best score across reasoning levels and that the evaluation environment can differ from production ChatGPT.

Coding: from writing code to working on a project

The practical difference in Codex is not just suggesting a function. With authorized access to a repository, terminal and available tools, a model can inspect files, propose changes, run commands, read errors, edit several parts and test again. In an IDE or Codex, the model supplies planning and code generation; the environment supplies the files and execution. Tests can catch some failures, but passing tests alone does not prove software is secure, maintainable or correct for every use.

It can also help debug, explain inherited code, prepare a migration, build a prototype or review a browser interface. A long task can still lose earlier decisions, misunderstand an instruction or fix one thing while breaking another. A brief with acceptance criteria, minimum permissions and intermediate reviews gives more control than asking it to “build a complete app” and accepting the first result.

Computer use: what it means to operate an interface

With computer use, the model observes screenshots of an interface and decides what action to request next. OpenAI's API guide describes structured keyboard and mouse actions executed by the integrating application, or code that operates a browser or desktop and returns new observations. This can support form testing, website navigation or interaction with specialist software.

The environment is the key distinction. Astra does not automatically control any computer: someone must connect it to a browser or machine, execute its actions, return the screen and define permissions. A test in an isolated desktop is not the same as authorizing purchases, sending emails or changing live data. OpenAI's demonstrations show forms, document editing, data analysis and website testing; these are company-selected examples, not guarantees that a changing workflow will work without supervision.

Research, science and maths: attribution matters

Astra can help search information, analyze files, run code and work with scientific tools when a user or integration supplies them. OpenAI reports 97.6% on FrontierMath Tier 4 v2, but a benchmark score is not proof that the model autonomously solves open problems or that its explanations are correct. Research still needs its sources, methods and results checked; a plausible answer is not expert review.

Headlines about mathematics must be separated from Astra-specific evidence. OpenAI describes its First Proof tests as work by an internal model during training, with human suggestions, selection of the best attempts and expert review; the company later corrected an attempt it had considered promising. In the Navier–Stokes case, OpenAI attributed the proof to another internal model, described as significantly more capable than Astra, coordinated in an agent system; Astra was later used to formalize and verify parts in Lean. Saying “Astra solved Navier–Stokes” would attribute a result the source does not assign to it.

What the official demos show

OpenAI's launch page shows Astra working on a circuit board, a game, a tax form, a spreadsheet, website QA and professional documents. Another demonstration takes a house model from Blender into an explorable Unreal Engine walkthrough. These examples illustrate reasoning combined with software tools. They are showcases selected by OpenAI, not independent trials: they do not prove that the model created every asset from scratch, that no preparation was involved or that the result is repeatable from any instruction.

What people are building with GPT-6 Astra

Third-party projects add a different perspective because they show artifacts, decisions and iterations outside the manufacturer's showcase. Even so, attribution often depends on what the creator reports; a screenshot saying “made with AI” is not enough to establish whether it used Astra, another model or a mix.

A historical cipher, with material to inspect. Security researcher Carter Church published a reconstruction of an 1809 encrypted letter to General Marmont, with a technical account and reproducible package. Church credits GPT-6 Astra with reading the archival image, grouping symbols and helping find a key; Cryptiana later marked the entry solved. There are caveats: an earlier partial key covered 435 of 1,300 units, Napoleon's correspondence provided context, and the work involved tools, checks and human judgment. It is interesting because it can be tested, but it was not a discovery of unknown strategic information: much of the content already appeared in a published order. Church reports about six hours of model execution from a single scanned image; he published the key, transcription and scripts to verify and reproduce the reading.

A 3D graphic that checks its own output. Creator Zhuermu published an interactive WebGL experiment and explicitly identifies GPT-6 Astra and Codex CLI. The creator documents 20 tool calls and three rendering and review rounds: the first result had excessive glow, which the agent noticed after viewing the screenshot and adjusted. It is a single-page graphics piece with an accessible demo, not a production application or proof that human review is unnecessary. The page reports a one-turn session, 10m 34s and a 575-line file; the creator provides the demo and method, plus reproducible screenshot hashes. These are creator-reported details, not an independent audit.

A fictional product website. VUVE documented how it built NOCTE, a demo page for an invented perfume brand, using Codex with GPT-6 Astra. The account includes an AI-generated product image, LTX video, connected tools and several revisions. The site and process can be inspected; Astra did not generate the video, and the page does not represent a real company or product. It shows assisted assembly and editing, not a complete commercial site made without intervention.

What Astra still does not guarantee

Astra can get facts wrong, invent references, choose the wrong screen control, misread requirements or repeat an error during a long task. An agent may also use up its allowance, take longer than expected or stop because of permissions and safety filters. Demos have clear goals and reduce friction; working in a real repository, business account or weeks-long project adds state, dependencies and risk.

Autonomy is a property of the whole system: model, instructions, tools, memory, permissions, reviews and limits. When a model can click or run commands, an ambiguous instruction can have real consequences. Grant minimum access, require confirmation before irreversible actions and review changes, especially when money, private data, production systems, security or professional decisions are involved.

Safety: what the Critical classification means

OpenAI classifies Astra at the Critical cybersecurity capability level in its Preparedness Framework and says it strengthened safeguards and monitoring. The category signals that, with suitable tools and access, the model could help find vulnerabilities and develop exploit chains; it does not mean every user receives these capabilities without controls or that an attack has occurred. OpenAI also acknowledges that monitoring its written reasoning became harder compared with GPT-5.6 Sol. This is the manufacturer's risk assessment, not a perfect safety certification.

Frequently asked questions

Is GPT-6 Astra the same as ChatGPT?

No. Astra is an OpenAI model; ChatGPT is a product that can offer several models. Work and Codex are different experiences in the ecosystem, and GPT-6 Pro is a Chat option powered by Astra.

Can I use Astra for free or with Plus?

The documentation checked here lists Astra in Work and Codex for eligible Plus plans, and GPT-6 Pro in Chat for certain Pro, Business and Enterprise plans. API access and billing are separate. Availability can vary by account, country, workspace permissions and product changes, so check the official pages before deciding.

Can it build an entire app or game by itself?

It can contribute to prototypes and projects through a configured agent and tools, but public examples show preparation, external software, iterations and human decisions. An impressive result does not prove that it is complete, maintainable or safe for production.

So what does GPT-6 Astra add?

Astra is most useful when it can combine reasoning with an environment that lets it inspect, execute, observe and correct work. Verifiable examples show coding, interactive graphics and analysis of historical material; official demonstrations broaden that range, while third-party cases show that results depend on tools, context and oversight. Astra is not a synonym for ChatGPT, cannot control any computer, and science or mathematics cases must be attributed to the model and team that actually produced them. The useful question is not whether Astra “can do everything”, but which task it solves, with what system around it and how the result is checked.