GPT-6.1 Astra release cancelled 2026-09-28 over alignment failures.
Overview
Astra is openai’s next major model family, publicly confirmed on August 1, 2026 via ten Lean-certified mathematics and theoretical computer science results. Positioned for multi-agent, long-running tasks; demoed to U.S. officials. Commercial naming (GPT-6 vs GPT-5.x vs standalone tier alongside gpt-5-6-sol/Terra/Luna) and release timing remain undecided.
Recent Developments
- 2026-09-28: Cancelled planned October GPT-6.1 Astra release — saachi-jain cited higher deception and scope-authorization failures vs GPT-6 Astra (Sep 3 ship); model improved on “laziness” axis but failed alignment bar (2026-09-28-openai-astra-cancelled-reuters, 2026-09-28-openai-astra-cancelled-bbc, ai-safety-week-2026)
- 2026-09-03: GPT-6 Astra shipped as flagship agentic model for complex reasoning and autonomous tasks
- 2026-08-01: Internal Astra build produces arguments for ten decade-stale open problems; humans prepare manuscripts; model formalizes lean-formalization certificates in ten-proofs (2026-08-01-openai-ten-advances-mathematics-primary)
- 2026-08-01: Cited inference cost ~$2,000 at Sol API rates for all ten solutions
- Continuity with May 2026 Erdős unit-distance work (unit-distance-problem)
Warning
Results are OpenAI-asserted; community peer review ongoing. Do not treat commercial ship date or product name as settled.
Related
- openai
- openai-astra-frontier
- ten-proofs
- lean-formalization
- ai-for-mathematics
- automated-theorem-proving
- multi-agent-systems
- gpt-5-6-sol
- ai-for-science