Deepa Kowalski

Posted on Sep 5

GPT-6 Astra Explained (2026): Access, Pricing, API & What Changed

#ai #openai #gpt6 #llm

GPT-6 Astra is OpenAI's new flagship model for complex work involving software, browsers, computer interfaces, and research. Announced on September 3, 2026, it began with a limited organizational rollout, with broader access planned over the following days. A launch announcement therefore does not establish that every eligible account already has access. OpenAI announcement

This explainer is for developers and AI practitioners deciding where to try Astra, how to budget API usage, and what to change in their workflows. It separates published specifications from vendor performance claims; it does not report PromptZone hands-on testing.

Verified September 5, 2026. For surrounding launches and updates, see the AI model releases timeline.

How to access GPT-6 Astra

ChatGPT plans: check both your plan and the experience

OpenAI's current help documentation makes a distinction that the launch announcement's broad plan list does not explain:

Plus: Astra is rolling out in ChatGPT Work and Codex.
Pro: Astra access includes Chat, Work, and Codex as rollout reaches the account. In Chat, the Astra-powered option is called GPT-6 Pro.
Business and Enterprise: GPT-6 Pro is rolling out in Chat; workspace permissions also affect access. Astra availability in agent experiences remains subject to rollout.
Free and Go: these are not listed as Astra rollout plans.

Do not assume that seeing a model in one experience means it is enabled in another. The help page explicitly says availability can differ between Chat, Work, and Codex. ChatGPT model availability

Once enabled, Astra uses the existing Work and Codex allowance. Plus and Business Standard include limited Astra usage; Pro and Business Premium can use their full existing allowance. Additional credits extend eligible usage, but purchasing credits does not move an account ahead in the rollout. Signing into Codex with ChatGPT uses plan billing; using an API key uses API billing. Work and Codex access and usage

Codex CLI: update before diagnosing model access

Use Codex CLI 0.153.3 or newer as the practical baseline for this walkthrough. For a fresh installation, choose 0.153.4, whose release notes explicitly fix Astra's visibility in the bundled model picker. Version 0.153.3 added Astra to the Amazon Bedrock catalogs and corrected clarification-tool guidance. Official Codex changelog

For an npm installation, update with:

npm install -g @openai/[email protected]

Then open your project directory, run codex, sign in, and use /model to select an available model. These are the documented CLI setup and selection steps. Codex CLI documentation

If an older installation displays “requires a newer version of Codex”, update the client before changing prompts. One qualification matters: OpenAI's help page currently states 0.153.0 or newer as the minimum. The ≥0.153.3 recommendation here incorporates subsequent fixes; it is not evidence that every earlier patch rejects Astra. Updating also does not establish account eligibility. Published minimum version

Microsoft Foundry

Microsoft's September 3 announcement says GPT-6 Astra is generally available in Microsoft Foundry. It lists Standard and Provisioned Throughput deployments in Global and US Data Zone geographies. This is Microsoft's availability statement for its service; OpenAI's account rollout remains a separate question.

Start from the Foundry model catalog linked in Microsoft's announcement. Standard offers usage-based deployment, while Provisioned Throughput provides reserved processing capacity. The service also provides identity, access controls, private networking options, and monitoring. Microsoft Foundry announcement

GPT-6 Astra API: model ID and first request

The published model ID is gpt-6-astra. OpenAI's developer quickstart uses it with the Responses API. Create an API key in the dashboard and make it available to your application through OPENAI_API_KEY; the official SDK reads that environment variable. API quickstart

After installing the Python SDK with pip install openai, this illustrative example requests a short planning response. It is a documentation-based example, not a request tested for this article:

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-6-astra",
    input="Outline a small web application's accessibility review in five steps.",
)
print(response.output_text)

The example produces text; it does not connect a browser or inspect a repository. Those actions require a tool integration. OpenAI's computer-use documentation explains that your application supplies the environment, executes requested actions, and returns observations. Selecting Astra alone does not create that environment. Computer-use integration

GPT-6 Astra pricing: the published rates

OpenAI publishes Standard API prices in US dollars per million tokens: $10 input, $1 cached input, $12.50 cache writes, and $50 output. For long-context requests, the published rates are $20 input, $2 cached input, $25 cache writes, and $75 output. API pricing

The long-context threshold is more than 272,000 input tokens, and the higher rates apply to the full request. A large context window consequently does not mean every request within that window has the same unit price. Astra model specifications

Fast mode costs twice the applicable Standard rates. Batch and Flex are listed at half the Standard rates. Fast mode is unavailable for Astra requests using EU data residency. These processing choices need to be part of a cost estimate, alongside any applicable tool charges. Processing prices and restrictions

Microsoft separately lists Standard Global short-context pricing at $10 input and $50 output, while Standard US Data Zone is $11 input and $55 output per million tokens. Use the rate for your selected deployment when estimating Foundry costs. Microsoft's published pricing

Teams budgeting separately rented infrastructure can also consult our cloud GPU pricing reference.

Context window and input limits

The Astra API model specification lists a 1,050,000-token context window, a 922,000-token maximum input, and 128,000 maximum output tokens. It supports text and image inputs, with text output, and lists an April 30, 2026 knowledge cutoff. Astra model specifications

These are API specifications. Use the published input and output limits separately when designing a request; the context headline is not an allowance for that many input tokens plus an unlimited answer. Also separate the knowledge cutoff from any information retrieved through tools.

The context size itself is not the generational change: the published Sol, Terra, and Luna specifications list the same context window and maximum output size. Sol specifications, Terra specifications, Luna specifications

What changed in computer use, browsing, coding, and cybersecurity?

Computer use

Computer use means operating software through its interface: reading screenshots, choosing actions, and checking subsequent results. OpenAI recommends a code-execution integration for Astra, while continuing to support the structured computer tool. The application remains responsible for running the browser or desktop environment. Computer-use documentation

OpenAI reports 72.6% for Astra versus 65.7% for Sol on its specified OSWorld 2.0 evaluation. Those are vendor-reported benchmark results, not a success-rate forecast for your application. Launch evaluation

Browsing

The system card emphasizes stronger resistance to prompt injection during browsing and professional computer work. OpenAI reports fewer unauthorized or potentially destructive actions than with Sol in the evaluated environments. This concerns how the agent behaves while acting on retrieved content, as well as the quality of its answer. Astra system card

Software engineering

Microsoft describes workflows that reproduce bugs, investigate causes, propose fixes, and prepare changes for developer review. This makes repository investigation and verification central to the coding proposition. Foundry engineering scenarios

OpenAI also describes an experimental Codex feature that lets Astra retain notes across context windows and search earlier windows for details. Its announcement says the feature must be enabled in configuration before a planned broader default rollout. Codex context changes

Cybersecurity

The system card identifies Astra as OpenAI's first model at the Critical cybersecurity capability level under its Preparedness Framework. With appropriate tools and access, OpenAI says it can discover unknown vulnerabilities and develop exploits across well-protected systems without guidance at every step. That assessment explains the stronger access and monitoring controls accompanying deployment. Astra system card

One-table comparison: Astra versus Sol, Terra, and Luna

Prices below are Standard API input/output rates per million tokens for short-context requests. Each model link supplies the specifications for its row. The roles summarize OpenAI's positioning, not an independent ranking.

Model and source	API model ID	Context window	Input / output price	Published role
GPT-6 Astra	`gpt-6-astra`	1,050,000 tokens	$10 / $50	Most demanding work across reasoning, coding, research, and computer use
GPT-5.6 Sol	`gpt-5.6-sol`	1,050,000 tokens	$4 / $20	Flagship GPT-5.6 model for complex professional work
GPT-5.6 Terra	`gpt-5.6-terra`	1,050,000 tokens	$2 / $12	Balance intelligence and cost
GPT-5.6 Luna	`gpt-5.6-luna`	1,050,000 tokens	$0.20 / $1.20	Cost-sensitive workloads at high volume

Sol's listed price is promotional, available at least through November 21, 2026. The model pages also document long-context premiums and cache-write charges, which this compact table omits. Sol pricing details

For a workload already meeting its requirements on Terra or Luna, the table supplies no evidence that replacing every request with Astra is necessary. Its higher token prices make task selection a meaningful engineering decision. Compare completed results and actual consumption on your workload before committing a whole pipeline.

For alternatives outside OpenAI, continue with our 2026 coding-model comparison.

API workflow changes beyond the model name

Asynchronous tools let Astra continue independent work while your application runs a tool. You mark a function or custom tool with async: true and return the result against its original call_id. OpenAI does not take over your job execution or background queue. This is useful when one lookup can run while another part of the task proceeds. Async tool calling

Mid-turn steering lets users add requirements while Astra works through a Responses API WebSocket connection. GPT-5.6 and earlier models do not support this mechanism. A steering message queues new direction; it does not undo completed actions or cancel tools already running. For example, a revised scope can guide the continuation without erasing what the application has already received. Mid-turn steering

Safety gating and stopped tasks

The system card describes phased Trusted Access for Cyber, also called Daybreak access, for authorized defensive work. Organizations can apply, and individuals can verify their identity and request access. Sensitive capability access therefore has additional conditions beyond ordinary model selection. OpenAI also acknowledges weaker monitorability than Sol under adversarial evaluation conditions. Access and safety findings

For API builders, misalignment monitoring can interrupt a conversation when it detects a potential problem involving actions such as sensitive-data access or destructive changes. A flag is a reason to review the work; it does not by itself establish misconduct. Monitoring can also flag legitimate activity. Monitoring documentation

If the API returns misalignment_policy_violation, OpenAI instructs applications to stop dispatching further actions and avoid automatically retrying the blocked workflow. Preserve relevant records and have the responsible person review what happened. Because monitoring is asynchronous, stopping does not reverse actions already completed. These error-handling requirements belong in the integration, alongside normal tool-result handling. Stopped-request handling

What it means for prompt engineers

Astra-specific prompting guidance is already published. OpenAI's guide identifies several behaviors worth configuring explicitly:

Clarification: state when Astra should proceed with reasonable assumptions and when missing information requires a question.
Instruction files: audit skills and AGENTS.md for conflicting guidance; the model may follow those instructions more sensitively.
Writing: specify the desired length, format, and tone because its default responses can be detailed and heavily formatted.
Delegation: define when subagents should be used if your application supports them.
Verification: calibrate testing to the task so small changes do not trigger unnecessary repeated checks.

These are documented starting points, not a tested PromptZone prompt recipe. Use them when adapting examples from the prompt library. OpenAI's Astra prompting guidance

The migration guidance also requires attention to API parameters: tool calling needs Responses, Astra does not support none reasoning, and sampling parameters such as temperature and top_p must be removed. A prompt rewrite alone does not resolve those compatibility changes. Astra migration guidance

Frequently asked questions

Is GPT-6 Astra available to every paid ChatGPT user now?

The documented rollout is gradual. Plus access is through Work and Codex, while Astra-powered GPT-6 Pro in Chat is rolling out to Pro, Business, and Enterprise; workspace permissions can also affect access. Plan availability

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

The published API pages list the same 1,050,000-token context window for Astra, Sol, Terra, and Luna. Context size alone therefore does not distinguish the models in this comparison. Astra, Sol, Terra, Luna

Is GPT-6 Astra always cheaper per completed task than Sol?

That is not established for arbitrary workloads. The published Standard short-context token prices are higher for Astra, so a task-level comparison must account for the work performed and tokens consumed. API prices

Can a steering message undo an agent's previous action?

No. OpenAI documents steering as a way to change subsequent direction; it does not undo prior actions, rewrite delivered output, or cancel tools already started. Steering limitations

PromptZone - AI Prompts, Guides and Tools for Builders