OpenAI unveiled GPT-6 Astra on September 3, 2026. What users refer to as “ChatGPT 6 Astra” is not a separate product name; rather, in ChatGPT, it is primarily offered as an option called GPT-6 Pro, while in Work, Codex, and API, the model name GPT-6 Astra is used. Therefore, even if you don’t see ChatGPT 6 Astra in the model selector, the announcement is not incorrect.
The core of this announcement is not just a new chat model with slightly improved scores. Astra is a high-performance agent model designed to perform complex reasoning, coding, computer usage, research, and document creation from start to finish within a single workflow. Simultaneously, it is the first model that OpenAI has classified cyber capabilities as Critical in its Preparedness Framework. This is why access control, monitoring, and safety shutdowns have become even more important alongside its capabilities.
A scene where complex research and coding tasks are completed as documents and software deliverables through a single intelligent work engine.
Let’s clarify the names: The relationship between Astra and GPT-6 Pro
The API model ID is gpt-6-astra. OpenAI developer documentation describes it as the highest-performing model for the most challenging end-to-end tasks. On the ChatGPT chat screen, it will be sequentially provided as GPT-6 Pro for supported paid plans. In Work and Codex, the name Astra may appear directly, and the deployment timing may vary by product.
| Usage Location | Display or Model Name | Currently Announced Access Method |
| — | — | — |
| ChatGPT Chat | GPT-6 Pro | Sequentially deployed to Pro $100·$200, Business, Enterprise |
| ChatGPT Work | GPT-6 Astra | Sequentially deployed to Plus·Pro·Business·Enterprise, usage varies by plan |
| Codex | GPT-6 Astra | Sequentially deployed to eligible plans, requires Codex CLI 0.153.0 or higher |
| OpenAI API | gpt-6-astra |
Access based on usage tier and limits, Free API tier not supported |
| Azure·AWS Bedrock | GPT-6 Astra | Check each platform’s region, account, and release status separately |
Deployment is not a simultaneous rollout to all accounts. The availability of the model selector, ChatGPT Work, Codex, and API may appear differently for several days. For Enterprise, administrators must enable model access in the workspace, and it may be disabled by default upon release.
Astra targets long end-to-end tasks
While existing models could also create code and documents, the process of gathering information, making judgments, and verifying results across multiple applications often led to context breaks at each stage. Astra focuses on completing long tasks such as software development, research, presentations, and professional documents by combining deep reasoning and computer usage.
OpenAI also introduced an experimental feature in Codex where, instead of continuously compressing previous content into a single summary when the context is full, Astra maintains work notes and retrieves past context. This approach aims to reduce the problem of failed attempts and test results disappearing during the summarization process in long refactoring tasks. However, this feature is experimental, and its future default behavior may change.
Benchmark figures must be considered with their environment. OpenAI’s announcement states that Astra achieved 57.9% on Terminal-Bench 4.0 and 96.0% on GPQA Diamond. These are results under disclosed evaluation conditions and do not guarantee the same success rate in all real-world projects. Specifically, tool permissions, storage size, test quality, and request difficulty significantly influence actual productivity.
API Specifications and Pricing
According to the official model documentation, GPT-6 Astra supports a context window of 1,050,000 tokens and a maximum of 128,000 output tokens. The knowledge cutoff date is April 30, 2026. It can receive text and image inputs but does not support audio or video inputs.
Standard API pricing is $10 per 1 million input tokens, $1 per 1 million cached input tokens, $12.50 per 1 million cached write tokens, and $50 per 1 million output tokens. Requests exceeding 272,000 input tokens will incur double input/cache fees and 1.5 times output fees for the entire request. There is also guidance that Batch and Flex are 50% of Standard, and Fast mode is double the corresponding fee.
예시: 입력 1,000,000 토큰 + 출력 100,000 토큰
단순 표준 요금 = $10 + ($50 × 0.1) = $15This example is a simple calculation excluding long input surcharges for 272,000 tokens, cache writes, and tool call costs like web search and computer usage. Actual costs must be calculated based on input length, output, cache, and tool usage per request.
The API supports Responses API, Chat Completions, streaming, Function calling, and Structured outputs. Tools such as Web search, File search, Code interpreter, Hosted shell, Apply patch, Skills, Computer use, and MCP can also be used with the Responses API. Fine-tuning is not currently supported.
What to expect from ChatGPT plans
GPT-6 Pro in ChatGPT Chat will be sequentially provided to Pro $100·$200, Business, and Enterprise plans. Plus users may receive limited usage in ChatGPT Work and Codex while Astra is being deployed, but it should not be assumed that GPT-6 Pro in Chat will have the same conditions.
Astra shares existing Work·Codex usage limits and can consume limits faster than GPT-5.6 Sol. The number of available uses depends on task size, input/output length, inference intensity, and Fast mode, not a fixed number of messages. Purchasing additional credits will not accelerate sequential deployment.
To use Astra in Codex, CLI 0.153.0 or higher and the latest ChatGPT desktop app are required. If the model is not visible, before assuming an account issue, check the app/CLI version, plan, workspace administrator’s permission, and the phased deployment status.
The most important change: Critical cyber capabilities
OpenAI announced that Astra has reached the Critical cyber capability standard of its Preparedness Framework. This means that with appropriate tools and access permissions, it can find unknown vulnerabilities and develop exploit methods in well-protected systems without human guidance at every step.
This assessment is not equivalent to concluding “Astra is dangerous, don’t use it.” Rather, it means that while it enables advanced defensive research, stronger controls are needed to reduce misuse and malfunction. OpenAI explains that it applies monitoring to all tool-use reasoning and uses conservative refusal thresholds and additional checks for high-risk users and tasks.
From a user’s perspective, even normal long-duration tasks may be slowed down or temporarily paused by safety systems. In ChatGPT and Codex, users may be asked to review tasks, and in the API, requests may be interrupted. Instead of treating this as a simple error, users should re-verify the scope of approval, target assets, and the intended tool actions.
OpenAI reported that Astra is more resistant to jailbreak and prompt injection than GPT-5.6 Sol, but does not claim that safety is complete. It also disclosed a limitation: Astra’s improved control over its Chain of Thought increases the possibility of evading monitors under some adversarial conditions. Do not rely solely on the model’s refusals and internal monitors; maintain separate application authentication/authorization, least privilege, sandboxing, and audit logs.
Practical checklist before adoption
First, rather than setting Astra as the default model for all requests, prioritize it for paths that justify its high cost, such as long reasoning, complex code changes, and multi-app tasks. For simple classification and short answers, routing to a cheaper model is more reasonable.
Second, do not grant the model Tool and MCP permissions broader than user permissions. For irreversible actions like operational changes, external transfers, deletions, and payments, implement server-side policies and human approval.
Third, reflect long input surcharges in cost alerts. A context of 1.05 million tokens does not mean every request should use the maximum. Reduce RAG search results, logs, and code scope, and validate caching strategies.
Fourth, run the same regression tests as for GPT-5.6 Sol. Do not just compare accuracy rates; also measure normal requests, prompt injection, out-of-scope data access, tool execution, and cost/latency.
Conclusion
GPT-6 Astra represents a broader change than just a new chat button in ChatGPT. It will be progressively offered as GPT-6 Pro in Chat, Astra in Work and Codex, and gpt-6-astra in the API. Its strengths lie in long end-to-end tasks and tool usage, but it also comes with significant costs and usage consumption, along with additional controls tailored to its Critical cyber capabilities.
Therefore, the decision to adopt should be based not on “Is it the strongest model?” but on “Does this task require Astra’s deep reasoning and computer usage, and can I control its permissions and costs?” The model has become more powerful, but this does not mean that final authority and safety responsibility can be handed over to the model.
Leave a Reply