Back to glossary

Term

GPT Astra

GPT-6 Astra is OpenAIs flagship from September 2026 — the first model at the Critical Preparedness tier, focused on computer use and cybersecurity.

GPT Astra — explained in more detail

GPT-6 Astra is the top model of OpenAIs GPT-6 generation. It started on September 3, 2026 in a limited rollout for participants of the Daybreak cybersecurity program and opened up gradually from September 4 for ChatGPT Plus, Pro, Business and Enterprise as well as through the API and AWS. Unlike the GPT-5.6 lines Sol, Terra and Luna, the focus is not text quality but computer use: Astra operates a machine on its own — forms, spreadsheets, navigating across web pages.

Key facts

  • Release: September 3, 2026 (limited, Daybreak), broader from September 4; proprietary access via the OpenAI API, ChatGPT plans and AWS.
  • Risk tier: the first OpenAI model at the Critical cybersecurity tier of the Preparedness Framework.
  • Two versions: the generally available one refuses advanced cybersecurity work; full capability stays behind the Daybreak application process.
  • Benchmarks per OpenAI: 100 percent on ExploitBench (GPT-5.6 Sol: 78.5 percent), 99.9 percent on ARC-AGI-3 with extended tooling (Sol: 7.8 percent).
  • Pricing per third-party analyses: around $10 per 1M input and $50 per 1M output tokens; context window above 1M tokens, knowledge cutoff April 2026.

Example / Practical use

The typical use case is not a prompt but a multi-step assignment: pull a dataset together from several web sources, transfer it into a CRM and operate the forms along the way. That class of task is where the gap to earlier models is largest. On general language ability the jump is small — the Artificial Analysis Intelligence Index puts Astra at 61.2, nearly level with Sol at 60.9.

Delimitation

Sol, Terra and Luna are the three price tiers of the GPT-5.6 family; Astra is not another tier within it but the next generation with a different focus. For copy, editorial work and research, Sol remains the better value. Astra pays off where a model is meant to act rather than answer. Note that Astra uses a reasoning technique that no longer records parts of its chain of thought as readable text — monitorability drops compared with Sol. No independent testing existed at the time of writing.

See everything in one place:GPT Astra