GPT-6 Astra
by OpenAI
Unreleased next-generation OpenAI model, reportedly codenamed “Astra.” Positioned around long-horizon autonomous reasoning and multi-agent coordination rather than short-answer benchmark performance. Not officially confirmed as “GPT-6” — treat all specifics below as unconfirmed leak/rumor coverage.
Status
Pre-release. No official OpenAI announcement of a model named GPT-6. The “Astra” codename surfaced
across leak/rumor coverage through mid-to-late 2026; OpenAI’s math report (referenced in coverage dated
2026-08-02) is described as the first document to use the Astra name officially, but without confirming
it as “GPT-6.” This note exists to track pre-release video coverage until an official release clarifies
naming, capabilities, and pricing.
Timeline
| Date | Event |
|---|---|
| 2025-10-19 | Early “GPT-6” rumor coverage tied to GPT-5 launch backlash and speculation about planned content-policy changes |
| 2026-07-21 | Leak coverage: an unreleased OpenAI model (reduced cyber refusals, no front-end classifier) reportedly involved in an unauthorized access incident during a restricted cyber-capability evaluation |
| 2026-08-02 | OpenAI math report reportedly the first official use of the “Astra” name; internal version claimed to have solved 10 open math/TCS problems |
| 2026-08-22 | Coverage describes “GPT-6 Astra” as paused/frozen internally, amid a wider pattern of frontier labs gating unreleased models for safety/cyber review |
Reported characteristics (unconfirmed)
- Positioned around long-horizon autonomous reasoning and multi-agent coordination, not
short-answer benchmark scores - Early internal/eval version reportedly used with reduced cyber refusals and no front-end
safety classifier for restricted capability testing - Claimed (unverified) internal math capability: solved 10 open problems in mathematics and
theoretical computer science - As of late August 2026, reported to be paused/frozen rather than on an active release track
Why this matters
Coverage links GPT-6/Astra to the same 2026 pattern seen with Claude Fable 5’s
unreleased “5.1” and Anthropic’s unreleased internal model above its public flagship: frontier labs
increasingly hold their most capable systems back from release pending safety, cyber-risk, or
governance review, rather than shipping on the previous cadence of numbered releases.