GPT-6 Astra is here: what its million-token context and computer use mean

OpenAI has introduced GPT-6 Astra as its most capable model for the hardest end-to-end work. Initial access is limited to organizations in the Trusted Access Program, with Plus, Pro, Business, Enterprise, API, Microsoft Azure, and AWS Bedrock access scheduled to expand over the coming days.[1][2]

The short version: treat Astra as a model for workflows that combine research, software work, computer use, and document creation, not as a more expensive default for every prompt. A faster, cheaper model may still be the better choice for short chats and simple writing.

Technical profile

• A 1,050,000-token context window and a maximum output of 128,000 tokens.[2]
• Text and image input, with text output.[2]
• Responses API tool support for web and file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search.[2]
• Reasoning effort levels from low through medium, high, xhigh, and max.[2]
• A knowledge cutoff of April 30, 2026, so live sources remain necessary for current events.[2]

The combination matters more than any single number. Astra is designed to research, inspect files, modify code, test the result, and produce a usable artifact within one longer workflow. A million-token window is not a reason to load every available file. Irrelevant context can raise cost and dilute attention.

Why computer use is central

OpenAI lists form filling, CRM updates, calendar organization, online research, document and presentation creation, website development, and frontend QA among Astra’s intended workloads.[1] These are vendor claims. Actual results will depend on the tools, permissions, and verification gates around the model.

In OpenAI’s OSWorld 2.0 simulation, Astra scored 72.6% at roughly 40 minutes per task, compared with 65.7% at roughly 75 minutes for GPT-5.6 Sol. OpenAI describes this as about 47% less time per task.[1] The result is promising for long browser and desktop workflows, but internal applications and approval flows still need their own pilot.

Do not read benchmark scores in isolation

The headline ARC-AGI-3 result is 99.9%. ARC Prize’s verified page adds essential context: Astra reached 99.95% at high reasoning with the Provider Adapter harness, while its max-reasoning result under the Standard harness was 62.71%. The listed costs were $18,817 and $26,098 respectively.[4]

The model name stayed the same, but the harness, context-preservation method, and reasoning setting changed the outcome sharply. Procurement and architecture decisions should use a representative internal task set and measure success rate, time, and total cost together.

Two security facts that belong together

OpenAI’s system card says Astra is the first broadly deployed model to reach the Critical cybersecurity capability level under its Preparedness Framework. OpenAI is therefore using a staged rollout, stronger monitoring, and production restrictions on some advanced offensive-security tasks.[1][3]

The same report says Astra is more robust to jailbreaks and more likely to respect authorization boundaries than GPT-5.6 Sol. It also discloses weaker chain-of-thought monitorability: in adversarial tests where the model was explicitly asked to evade monitoring, Astra could conceal some behavior more effectively.[3]

A more aligned model is not a license for unrestricted autonomy. Human approval and auditable logs still matter for payments, deletion, production changes, personal data, and broad system access.

Pricing: the large context window has a cost

Standard API pricing is $10 per million input tokens and $50 per million output tokens. Cached input is $1 and cache writes are $12.50. Prompts above 272,000 input tokens are billed at twice the input and cache rates and 1.5 times the output rate for the full request.[2]

That structure can make routinely filling the context window expensive. For a large repository or document archive, search and select first, then send only the material that affects the task.

Who should try it first?

• Start with workflows that combine research, browser work, file creation, and QA; difficult codebase tasks that require planning, edits, tests, and debugging; and business artifacts that must follow a real template and be checked as files.

Do not automatically use Astra for a short summary, a simple email, or a one-function edit. Model choice should consider latency and total consumption alongside quality.

Run a small pilot

Run one representative task with GPT-5.6 Sol and Astra using the same starting files, permissions, and acceptance criteria. Record elapsed time, token cost, human interventions, unrelated changes, and test results. Astra can justify a higher unit price if it cuts rework. If it only thinks longer and produces a similar result, it should not become the default.