GPT-6 Astra is a proprietary language model from OpenAI, released in September 2026. Its focus is on computer use and agentic work, meaning operating applications via the graphical interface and handling multi-step tasks over extended periods without ongoing supervision. According to OpenAI, the model achieved the best measured results at release in computer use, browser research, software development, cybersecurity, science, and specialized tasks. The context window holds 1.05 million tokens with up to 128,000 token output.
For enterprises, two points are critical. Processing can be limited to European data centers, but only in standard processing and only with an appropriate data retention agreement. At the same time, Astra is the first OpenAI model to reach the highest risk level of the in-house Preparedness Framework in cybersecurity, resulting in additional controls and access restrictions. Open model weights are not available. You use the model via OpenAI, the major cloud platforms, or an access point such as our MaaS Proxy.
GPT-6 Astra
OpenAI
September 2026
Proprietary (no open weights)
Proprietary, reasoning with five effort levels, input text and image, output text
Not published
Not published. Reasoning, controllable via five effort levels
1,050,000 tokens (max. 128,000 token output)
The standard model with API identifier gpt-6-astra. Available via the OpenAI API, Microsoft Azure and AWS Bedrock, as well as via our MaaS Proxy. Regional processing in Europe is only possible in this processing mode.
The same model with up to double the speed at double the price per token. Important for European customers: regional processing in Europe is not available in this operating mode.
An extended variant that OpenAI provides in the ChatGPT Pro, Business, and Enterprise plans. For integrating your own applications, API access is the relevant method.
GPT-6 Astra excels at operating applications, but it is neither inexpensive nor approved for every use case. We work with you to determine where deployment is worthwhile, which data protection requirements can be met, and where a different model is the better choice.
OpenAI publishes neither parameter count nor architecture and provides no information on training data or training procedures for GPT-6 Astra. The model’s knowledge extends to April 30, 2026. Fine-tuning with your own data is not provided for this model.
Reasoning, meaning the step-by-step thinking through a task before answering, is controlled via five effort levels from low to max. For long workflows, OpenAI has modified context management: in the Codex development environment, the model can, as an experimental feature at release, maintain notes across multiple context windows instead of compressing earlier sections into a single summary. Older context windows remain searchable. This is why OpenAI recommends the model for multi-hour tasks such as extensive refactoring.
Data protection: Processing in the EEA and Switzerland is possible in standard processing. OpenAI offers Zero Data Retention to eligible API customers.
Cybersecurity classification: Astra is the first OpenAI model at the highest level of the Preparedness Framework for cybersecurity. Advanced cybersecurity applications are initially restricted and are being gradually released through the Daybreak Blue program. Stricter limits apply to accounts with elevated risk.
Restrictions: According to OpenAI, the model’s reasoning processes are harder to monitor than its predecessor, and protective mechanisms may interrupt legitimate work.
GPT-6 Astra demonstrates its strengths where a model must operate applications independently and execute multi-step workflows without ongoing supervision. It is less suitable for large volumes of simple tasks such as classification, for workflows in the cybersecurity domain, and for scenarios requiring full control over infrastructure. In those cases, smaller or open models in your own data center are the better choice. We are happy to clarify this with you in an AI consultation.
Leading in computer operation: Significantly ahead of its predecessor in OSWorld and Agents’ Last Exam, with 47 percent less time per task.
Very large context window: 1.05 million input tokens and 128,000 output tokens.
Strong in science and mathematics: 97.6 percent in FrontierMath Tier 4 and 96.0 percent in GPQA Diamond.
Processing in Europe possible: In the EEA and Switzerland, in standard processing.
Fewer fabricated statements: According to OpenAI, 4.2 percent hallucinations compared to 12.2 percent in its predecessor.
No operation in your own data center: The model is proprietary, fine-tuning is not provided.
Restrictions in cybersecurity: Access is limited, protective mechanisms may interrupt legitimate work.
Significantly more expensive per token: Approximately two and a half times more expensive than Claude Opus 5.5, additionally more expensive from 272,000 input tokens.
Not leading in all tests: According to Anthropic, Opus 5.5 leads in Terminal-Bench 4.0, GDPval-AA, and Humanity’s Last Exam.
We integrate GPT-6 Astra into your systems via our MaaS Proxy, with centralized budget control, transparent billing in euros, and gateway operation in German data centers. On request, we evaluate which applications are better suited to open models in your own infrastructure.
OpenAI reports the greatest improvement over GPT-5.6 Sol in computer operation and agentic tasks. In AutomationBench, the score rises from 18.1 to 41.4 percent, in Terminal-Bench 4.0 from 37.3 to 57.9 percent, and in Terminal-Bench Science from 22.4 to 64.6 percent. Tasks in OSWorld are solved on average in 40 instead of 75 minutes.
The prerequisites are in place: processing can be limited to the European Economic Area and Switzerland, and OpenAI offers Zero Data Retention to eligible API customers. Both apply only to standard processing, not to Fast Mode, and require an appropriate agreement. The specific assessment depends on your use case, your data, and the chosen access method. We provide individual consultation on this.
OpenAI classifies Astra as the first in-house model in the highest risk level of its Preparedness Framework for cybersecurity. According to OpenAI, the model can identify unknown security vulnerabilities and independently develop ways to exploit them. Defensive use cases such as malware analysis are being gradually enabled through the Daybreak Blue program. Anticipate that protective mechanisms may also interrupt legitimate work.
The list price per token is approximately two and a half times higher than Claude Opus 5.5. From 272,000 input tokens, the input price doubles and the output price increases by half. Fast Mode costs double the base price. Via our MaaS Proxy, you bill in euros.
No. The model is proprietary and only usable via OpenAI, the connected cloud platforms, or an access point such as our MaaS Proxy. Fine-tuning with your own data is also not provided. Those who need full control over infrastructure are better served by open models, such as GLM-5 or DeepSeek-V4-Pro.
Would you like individual advice?
Our AI experts are here for you!
