Carregando conteúdo…
Artificial Intelligence
GPT-6 Astra arrives with a focus on agents, programming, research, and computer usage. See the new features, limits, pricing, and the comparison with the GPT-5.6 family.
By Ederson Andrade · September 10, 2026 · 9 min read
Audio
11 minutes, approximately
OpenAI announced on September 4, 2026 the GPT-6 Astra, the company's new flagship model for complex tasks. In the official presentation, the company calls it its smartest and most capable model yet, focusing on reasoning, programming, research, web browsing, computer use and document production.
The announcement goes beyond a new version of chatbot. The main bet is to transform the model into an agent capable of conducting long jobs that cross code, browser and professional programs. He can also receive a new instruction while working, continue reasoning during some tool calls and change the level of reasoning effort in the middle of the conversation.
Availability, however, begins gradually. According to OpenAI, initial access is intended for companies participating in the Trusted Access Program. The arrival of the API and the Plus, Pro, Business and Enterprise plans was promised for the days following the announcement. This means that the name can appear in official materials before it is released to all accounts.
GPT-6 Astra is OpenAI's new, most advanced general-purpose model. Its identifier in the API is gpt-6-astra, and the company recommends it for end-to-end jobs that require multiple steps, tools, and chained decisions.
In practice, this description includes tasks such as examining a code base, searching for documentation, operating interfaces, producing files, and reviewing the result. Rather than just answering an isolated question, the model is designed to maintain the purpose of a long task and react to changes made by the user during execution.
The word "smart" needs to be read with context. It is the classification used by OpenAI itself to position the product. It does not mean that the model is infallible, conscious, or better at any imaginable task. Models can still misinterpret data, produce false information, and perform a poor sequence of actions when given insufficient context.
Previously, an agent often had to stop and wait for the result of a tool to continue. With the new asynchronous mode, Astra can continue to reason, call other tools, or respond to independent parts of the request while an external operation is still in progress.
This can reduce idle time in long flows, but it does not deliver unrestricted autonomy to the model. The documentation itself clarifies that the application remains responsible for running the tool, controlling pending work, and returning the correct result using the caller ID.
The resource called mid-turn steering allows you to send a new orientation while the model is still working. In a WebSocket connection, the API preserves what has already been completed and incorporates the fix in continuation.
This is a relevant change for tasks that take several minutes. If the user notices that a report is following the wrong format or that a condition has changed, they can correct the route without necessarily restarting the entire process.
Astra accepts levels of reasoning low, medium, high, xhigh e max. An application can augment this effort in the face of a difficult step or reduce it to a simple continuation while preserving the prefix used by the cache.
Unlike GPT-5.6 models, Astra doesn't offer the level none. Therefore, even at the lowest setting, it retains some reasoning processing. This decision suits its positioning for difficult tasks, but it can make it unsuitable for very simple and cost- or latency-sensitive jobs.
OpenAI claims that Astra is also its most aligned model yet. The company has added a system that asynchronously monitors for signs of misalignment and can generate security alerts.
This is a manufacturer's assertion, not a zero-risk guarantee. Security continues to depend on the permissions granted to the agent, the validation of inputs, the limits of the tools, and human confirmation before major actions.
The comparison below uses the specifications and prices published by OpenAI as of September 4, 2026. Amounts are charged per 1 million tokens on the API, before any additional fees for tools or processing modes.
| Model | Recommended use by OpenAI | Background | Maximum output | Knowledge Cut | Entry | Cached Entry | Output |
|---|---|---|---|---|---|---|---|
| GPT-6 Astra | Complex end-to-end jobs | 1.05 million | 128k | April 30, 2026 | $10 | $1 | US$ 50 |
| GPT-5.6 Sun | Complex professional work | 1.05 million | 128k | February 16, 2026 | $4 | $0.40 | US$ 20 |
| GPT-5.6 Earth | Balance between capacity and cost | 1.05 million | 128k | February 16, 2026 | US$ 2 | $0.20 | $12 |
| GPT-5.6 Luna | High volume at reduced cost | 1.05 million | 128k | February 16, 2026 | $0.20 | $0.02 | $1.20 |
The table shows that the context window has not increased relative to the GPT-5.6 family. All four models can receive up to 1.05 million tokens and produce a maximum of 128k output tokens. The change is mainly in the declared capacity, behavior during long tasks, knowledge cut, and price.
Astra costs two and a half times the value of GPT-5.6 Sol per input token as well as per output token. In a simple example, 1 million input and 1 million output tokens would cost $60 on Astra, versus $24 on Sol, without considering caching, tools, and other billing modes.
OpenAI says that Astra achieves better results by using substantially fewer output tokens in various evaluations. If this holds true in a real application, the cost per task may grow less than the difference in the table suggests. Still, each company will need to test its own flow. Price per token, price per task, and result value are different measures.
The model accepts text and images as input, but the direct output is textual. Audio and video do not appear as native Astra modalities in the datasheet. It can use imaging tools or other services within a stream, but this is different from producing these modalities directly from the model.
The context window of 1.05 million tokens allows you to work with large sets of documents or codebases, but it does not eliminate the need to select relevant information. Putting more content in the prompt can increase cost, latency, and distractions. The price list itself informs that requests with more than 272 thousand input tokens receive higher charge multipliers.
The knowledge cutoff is April 30, 2026. This does not mean that Astra automatically knows what happened after that date. For recent information, it relies on tools such as web search and the data provided by the application.
In the Responses API, Astra supports web search, file search, image generation, code interpretation, hosted shell, applying changes to files, skills, computer usage, MCP, and tool discovery.
It also supports streaming, function calls, and structured outputs. Fine-tuning is not available. For developers, another important detail is that tool calls must use the Responses API, although the model also accepts conversations through the Chat Completions endpoint.
Migration is not just about changing the name of the model. OpenAI advises removing parameters such as temperature, top_p e top_logprobs, which are not supported in Astra. Applications that used reasoning none need to start at low and compare behavior.
Not at the time of the announcement. OpenAI described a phased release. First come companies in the trusted access program; API and paid plans appear in the announced sequence. The documentation also informs that the free tier of the API does not support the model.
It's important to separate three things: the GPT-6 Astra model, the API used by developers, and the ChatGPT product used by the public. They relate, but availability, tools, and limits may change depending on the plan and region.
There's also no reason for every application to migrate immediately. Sol remains cheaper for complex professional work. Terra can handle tasks that need to balance quality and budget. Luna costs a fraction of Astra and makes more sense in sorting, extracting, and high-volume repetitive operations.
The most interesting point of GPT-6 Astra isn't just about generating a better response. It's about closing the distance between explaining a job and getting it done. Asynchronous calls, run-in-the-run correction, and adjustable reasoning all point to agents staying active longer and handling multiple steps without losing context entirely.
This direction also increases the accountability of those who build the systems. A mistake in a sentence is different from a mistake made by an agent with access to files, browser, terminal, and external services. The greater the ability to act, the more important scope, authorization, audit logs, and clear boundaries become.
The same movement appears outside the computer. Products such as Zeehoo AIPAL Dock they try to bring persistent assistants to physical objects. Astra tackles another part of the problem: coordinating digital tools and longer professional jobs.
It is correct to say that OpenAI presents the GPT-6 Astra in this way. The official documentation claims high-end performance in computer use, navigation, software engineering, science, and professional work. It also says that the model has outperformed previous versions in different evaluations with fewer output tokens.
What would not be correct is to convert this message into a universal percentage without the detailed results, test conditions, and independent comparisons. There is no single test that represents programming, research, creativity, factual accuracy, security, and speed at the same time.
For now, the confirmed picture is this: Astra occupies the top of OpenAI's catalog, brings new features to agent flows, maintains the large context window of the 5.6 family, updates the knowledge cut, and charges a premium price. Its real advantage will depend on testing on concrete tasks, especially when the cost and risk of automatic actions come into account.