GPT-6 Astra: What's New, How Much It Costs, and Who Can Use It

OpenAI launched GPT-6 Astra on September 3, 2026 as its new state-of-the-art model for complex end-to-end tasks. The innovation isn’t just about “responding better”: Astra is designed for long workflows that combine reasoning, programming, navigation, tool usage, computer use, research, and document creation. The rollout, however, is gradual: OpenAI notes that it is starting with companies in its Trusted Access Program and that access via API and for the Plus, Pro, Business, and Enterprise plans will become available over the next few days.

For a business, the useful question isn't whether Astra is “the best model,” but rather in which tasks its greater capabilities can offset a token price that is significantly higher than that of GPT-5.6 Sol. It's also important to review API changes, security controls, and the effective cost per task before migrating a workflow that's already up and running.

In a nutshell

• GPT-6 Astra is OpenAI's new state-of-the-art model for complex reasoning, coding, research, computer use, and multi-step agentive workflows.
• The rollout is gradual: it begins with companies in the Trusted Access Program; OpenAI will announce the API and access for Plus, Pro, Business, and Enterprise in the coming days.
• On API, it costs US$10 per million input tokens and US$50 per million output tokens; GPT-5.6 Sol costs US$4 and US$20, respectively.
• It maintains a context window of 1.05 million tokens and a maximum output of 128,000 tokens.
• It introduces features such as asynchronous calls to tools, instructions during execution, and changes in reasoning level without losing the cached prefix.
• It requires no effort to reason it out none; OpenAI recommends the Responses API for tool calling.
• OpenAI classifies Astra as "Critical" in terms of cybersecurity capabilities under its Preparedness Framework, so the launch incorporates enhanced controls and monitoring.
• It is not advisable to automatically migrate all processes: routine or cost-sensitive tasks may still be best handled using more economical models.

What Is GPT-6 Astra, and How Does It Differ from GPT-5.6?

OpenAI introduces GPT-6 Astra as its most capable model for end-to-end tasks. That description is important because the focus of the launch is not just on a single conversation, but on tasks that require maintaining goals, using various tools, course-correcting, and completing a sequence of actions.

According to official documentation, Astra is specifically designed for computer use, web browsing, software engineering, science, and professional work. It also retains capabilities that were already present in GPT-5.6, such as Structured Outputs, streaming, programmatic use of tools, multi-agent orchestration, prompt caching, persistent reasoning, compaction, and computer use.

The practical difference becomes apparent when a task ceases to be a simple “question → answer” and becomes a process: reviewing information, deciding which tool to use, waiting for results, continuing while another tool is working, incorporating a user’s correction, and producing a verifiable final result.

Who can use GPT-6 Astra, and when will it be available?

The launch should not be interpreted as meaning that the feature is immediately available to all accounts. OpenAI documents a phased rollout.

• Start of deployment: companies participating in the Trusted Access Program.
• API: OpenAI says that access will be expanded over the next few days.
• ChatGPT: The documentation mentions the Plus, Pro, Business, and Enterprise plans as part of the rollout announced for the coming days.
• Free: The documentation reviewed at the time of this review does not include it among the plans announced for Astra.

GPT-6 Astra Specifications and Price

At the standard rate, Astra costs 2.5 times more per input and output token than GPT-5.6 Sol. That difference makes a comparison per million tokens necessary, but not sufficient.

OpenAI states that, in several evaluations, Astra achieves better results using substantially fewer output tokens, so some tasks may end up with a lower estimated cost per result even with a higher unit rate. This should be treated as a hypothesis that each company must evaluate: the savings depend on the task, the number of retries, the reasoning used, and how much human correction is still required.

The long-term context can affect costs more than it seems

The 1.05 million-token window allows you to work with large amounts of context, but the Astra documentation sets forth an important condition: Requests with more than 272,000 input tokens are subject to higher fees for the entire request. OpenAI specifies 2× for input and cache and 1.5× for output for those prompts.

Therefore, “supports one million tokens” does not mean that it is always advisable to send one million. In a real-world application, it is best to assess how much context the model needs, retrieve only the relevant documents, and take advantage of the cache when a stable prefix exists.

For repetitive jobs with lengthy instructions, caching costs US$1 per million tokens on Astra. Cache writes are billed at 1.25× the non-cached rate. Batch and Flex offer a 50 % discount compared to Standard, while Fast is billed at twice the applicable rate; additionally, OpenAI notes that Fast is not available for Astra with data residency in the European Union.

New Astra Features in the API

Part of the release involves how an application can guide the model as it works. OpenAI documents three changes that are particularly relevant for agents and long-running processes.

Asynchronous Calls to Tools

A function or tool can run asynchronously while Astra continues to reason, handle another part of the request, or use other tools. This reduces the need to block the entire execution while a slow integration completes.

For an enterprise application, this feature is useful in situations where a query to an external service takes several seconds or minutes: the agent can continue working independently and receive the result later using the call ID.

Changes to Instructions During Execution

With mid-turn steering, an application can send a correction or additional requirement while Astra is still working. Through a WebSocket connection, the model retains the work already completed and incorporates the new instruction as it continues.

This is useful in workflows where the user is monitoring a long-running task and needs to adjust the scope without starting from scratch. It does not eliminate the need to design states, permissions, and auditing; it simply makes the interaction more flexible during execution.

Changing the level of reasoning without reconstructing the context

The API allows you to add a configuration_update to increase or decrease the cognitive load during a conversation by keeping the prefix in cache. An application can reserve high levels for difficult steps and reduce them for routine tasks.

Astra admits low, medium, high, xhigh y max. Unlike GPT-5.6 Sol, it does not offer none.

What tools does it support?

The model's official documentation lists support via the Responses API for tools such as web search, file search, image generation, Code Interpreter, hosted shell, Apply Patch, Skills, computer use, MCP, and tool search. It also supports function calling and Structured Outputs.

This does not mean that all of them must be enabled at the same time. In a business agent, More tools mean more permissions to manage, more potential errors, and more actions that need to be auditable. The selection should be based on the principle of minimal access.

What to Check Before Migrating from GPT-5.6

The basic migration begins by changing the identifier to gpt-6-astra, but OpenAI documents some differences that you should review before making the change in production.

1. Use the Responses API for tool calls: Astra supports Chat Completions, but the official guide specifies Responses for flows that use tools.
2. Review the reasoning process: If your application uses none o minimal, OpenAI recommends starting with low and measure.
3. Remove unsupported parameters: The guide instructs you to remove temperature, top_p y top_logprobs; in Chat Completions as well logprobs.
4. Try the existing instructions: OpenAI notes that Astra is more sensitive to instructions found in skills and files such as AGENTS.md. It is a good idea to audit these resources before granting them access to production.
5. Recalculate costs: Don't just multiply your historical token counts by the new rate; measure tokens, retries, quality, and human time using the actual model.
6. Check permissions and tools: A more capable model does not automatically justify expanding the scope of the credentials it receives.

If you're coming from GPT-5.6 Sol, you can also check out the UNEMPRENDE article on The launch and features of GPT-5.6 to determine which part of the change is due to the model and which part is due to your application's architecture.

When Might It Make Sense to Use Astra?

Agents that perform long, multi-step tasks

Astra makes the most sense when the outcome depends on pursuing a goal through several steps: researching, using tools, reviewing results, correcting errors, and continuing. If the task can be resolved with a short, verifiable answer, that additional capability may not justify the cost.

Software Development and Maintenance

OpenAI positions Astra for software engineering and computer-aided work. The potential benefit lies in workflows where the model must understand a project, modify files, use tools, test changes, and respond to intermediate results. In these cases, the useful metric is not how many lines of code it produces, but how many tasks are completed correctly with less human rework.

Research and professional work involving many documents

The expanded context window and multi-stage workflow can be useful for document review, professional analysis, and source synthesis, provided that traceability is maintained. A well-written report does not replace the verification of figures, contracts, regulations, or specialized decisions.

Processes that combine a browser and professional software

When a task requires navigation, interaction with applications, and maintaining context across tools, improvements in computer use and steering may be more important than a difference in a single benchmark. These workflows also require permission controls, action logs, and error recovery.

What Do the First Cases Published by OpenAI Reveal?

On the day of the launch, OpenAI published examples of companies that tested Astra. These serve as indicators of the type of work the model is prioritizing, but should not be interpreted as guaranteed results for another company.

• Legora: It reported that an agent using Astra reviewed 41 documents in a matter of minutes as part of a financial statement reconciliation workflow and found all four errors that had been intentionally introduced for the test. In its specific benchmark for that workflow, Legora reported an improvement of nearly 40 % compared to the previous model.
• Playco: It reported 50 % fewer manual corrections in its game prototyping process compared to the previous model and highlighted improvements in spatial reasoning, reference recreation, and validation within the engine.

Both cases follow a common pattern: they are not simple chats. They are processes in which the model receives context, uses tools, and must produce a verifiable result. That is a more useful indicator for deciding on a test than directly extrapolating from the published percentages.

When Automatic Migration Is Not Recommended

GPT-6 Astra does not render the more affordable models obsolete. OpenAI continues to offer GPT-5.6 Sol, Terra, and Luna in its product lineup, each with different prices and profiles.

The decision should take into account the total cost of a task: tokens, tools, execution time, failures, retries, and minutes of human review. A more expensive model can be cost-effective if it sufficiently reduces rework; a cheaper one may still be better if the task is simple and stable.

Why the Astra launch has different safety measures

OpenAI classifies GPT-6 Astra as its first model to reach the level Critical cybersecurity capabilities under its Preparedness Framework. The company explains that, with the right tools and access, the model can conduct vulnerability research at a level that requires enhanced protections prior to a broad deployment.

For the launch, OpenAI outlines additional measures for isolation, access control, movement tracking, alignment testing, and restrictions on certain dual-use capabilities. It also maintains verified access programs for advanced work in cybersecurity and biology.

For a typical company, the practical lesson is not to try to use the most sensitive capabilities, but to recognize that A more capable agent requires authorization controls that are at least as rigorous as the process it automates. If you connect email, repositories, browsers, files, or internal systems, review minimum permissions, logs, and actions that require human approval.

To further explore this topic, check out the UNEMPRENDE guide on Privacy and Sensitive Data When Using AI in a Business.

How to Evaluate Astra Before Putting It into Production

Don't switch to a new model in production based solely on the advertisement. Prepare a set of tasks that reflect the actual work and compare it to the model you're already using.

1. Select between 20 and 50 representative tasks: It includes easy, difficult, and ambiguous cases.
2. Freeze the definition of success: Decide which answer counts as correct before looking at the answer key.
3. Calculate total cost: input and output tokens, tools, retries, and review time.
4. Final measurement: It tracks how many tasks achieve the expected result without manual intervention.
5. Measures tool errors: It is not enough to evaluate the written response if the agent is required to take action.
6. Compare levels of reasoning: Do not use max by default; it identifies the minimum level required to maintain the necessary quality.
7. Test limits and permissions: This includes situations in which the agent must stop, request authorization, or refuse to take an action that is beyond their scope.
8. Keep rollback: Keep the previous settings until you verify that the change is stable.

Metrics That Help You Make Decisions

• Average cost per task completed correctly.
• Percentage of tasks completed without additional human intervention.
• Review time by result.
• Number of retries or corrections.
• Total latency of the flow, not just the first token.
• Errors in using tools or interpreting their results.
• Permission issues, out-of-scope instructions, or unauthorized actions.
• Output quality evaluated using the same criteria as for the previous model.

The most useful metric is usually cost per acceptable result, not cost per million tokens. That comparison highlights the difference between paying more for a successful run and paying less for multiple runs that require correction.

Frequently Asked Questions

Is GPT-6 Astra now available to everyone?

No. OpenAI announced a phased rollout on September 3, 2026. It begins with companies in its Trusted Access Program, and the company states that the API and the Plus, Pro, Business, and Enterprise plans will gain access over the next few days. Actual availability should be checked for each account.

How much does GPT-6 Astra cost via the API?

The published standard rate is US$10 per million input tokens, US$1 per million cached inputs, and US$50 per million output tokens. Additional conditions apply for prompts longer than 272,000 tokens and for modes such as Batch, Flex, and Fast.

Is Astra more expensive than GPT-5.6 Sol?

Per token, yes: the standard input and output rate is 2.5 times that of GPT-5.6 Sol. OpenAI states that Astra may use fewer output tokens in certain evaluations, so the cost per task may vary. This needs to be measured based on actual usage.

Does Astra have a larger context window than GPT-5.6 Sol?

Not according to the current spec sheets: both list 1,050,000 context tokens and 128,000 maximum output tokens. The update focuses on capabilities, behavior, and new execution features rather than on increasing those two figures.

Should I replace GPT-5.6 Sol with Astra?

Not automatically. Astra makes sense when the increase in capacity reduces errors, rework, or human intervention enough to offset its cost. For simple or high-volume tasks, Sol, Terra, or Luna may still be more efficient options.

Does Astra work with Chat Completions?

The model documentation includes Chat Completions, but the migration guide recommends using the Responses API for tool calls. If your application relies on tools, you should treat Responses as the primary method.

What does it mean for Astra to have "Critical" cybersecurity capabilities?

This is a classification within OpenAI's Preparedness Framework that triggers enhanced security requirements. It does not mean that sensitive capabilities are made available without restrictions to any user. The deployment includes controls, monitoring, and verified access paths for certain advanced use cases.

Astra should be viewed as a driver for processes, not as an automatic update

GPT-6 Astra It represents a significant change for workflows in which a model must sustain a long-running task, coordinate tools, and adapt as it works. It also comes with a higher price per token and more stringent security controls.

For a company, the best way to adopt this approach is straightforward: select a few challenging processes, compare them to the current model, measure the cost per acceptable result, and scale up only where the difference is demonstrable. For simple tasks, migrating simply for the sake of novelty can increase expenses without producing a proportional benefit; for complex processes, a sufficient reduction in rework can completely change the equation.