Definition
Astra is OpenAI's named GPT tier introduced with GPT-6 and positioned for demanding reasoning, software engineering, research, and computer use. The documented API model is gpt-6-astra. Astra identifies a particular model offering; frontier model and reasoning model describe broader categories.
Origin
OpenAI announced GPT-6 Astra on September 3, 2026. Its launch report describes a model trained through pretraining and reinforcement learning, with evaluations across professional work and agent tasks. OpenAI is the developer; the release does not identify a sole inventor of the model.
Scope and capability claims
The model documentation lists text input and output, image input, and configurable reasoning effort. Computer use requires a surrounding interface and execution environment that let the model perceive and act on applications.
OpenAI positions Astra above Sol and Luna for demanding work. That is a vendor-defined tier and a task-dependent capability claim. It does not establish that Astra is the best route for every request or that an agent using Astra always respects its operating boundaries.
The launch report classifies Astra at OpenAI's Critical cybersecurity threshold. That threshold belongs to the company's Preparedness Framework; it is not a general definition of intelligence or an authorization to perform a security task. The report distinguishes evaluations without production safeguards from the deployed safeguard stack.
Operational significance
Preserve the model identifier, reasoning setting, harness, and tool access when interpreting reported results. A benchmark completed with broad tool access can answer a different question from the same task under production permissions.
Use workload evaluation to decide when a more costly route is justified. Verify consequential outputs and keep approval checks in the application even when the selected model has stronger results on alignment evaluations.
Distinguish it from nearby terms
- GPT is the model lineage; Astra is a named offering within it.
- Sol and Luna are separate GPT tiers, not lower reasoning settings on Astra.
- A frontier model describes a position near current capabilities across selected tasks.
- Computer use is a capability shared by models and systems from several developers.
Check your understanding
An Astra-based agent passes a coding benchmark with unrestricted tools. Which evidence would you need before allowing the same agent to change a production service under narrower permissions?