Omanghana
OpenAI has disclosed that an internal version of Astra, described as the company’s next major AI model, contributed to progress on several longstanding problems in mathematics and theoretical computer science.
However, the company has not officially announced Astra as a publicly available GPT-6 product or released pricing, access plans and benchmark results for the model.
In a research publication, OpenAI said the internal system helped produce results across 10 areas, including high-dimensional sphere packing, coding theory, arithmetic circuit complexity, quantum complexity and lattice cryptography. Human researchers subsequently prepared the arguments as manuscripts with assistance from the model, before the work was formalized through Lean certificates.
The disclosure suggests that OpenAI is testing Astra on advanced research tasks requiring sustained reasoning, mathematical exploration and formal verification.
Among the reported achievements were new bounds for high-dimensional sphere packing, progress on binary and spherical codes, a construction concerning non-sofic groups and results addressing established conjectures in geometry, complexity theory and extremal combinatorics.
Despite the research progress, OpenAI has not published official information confirming Astra’s training infrastructure, context-window capacity, API prices or availability through ChatGPT, Microsoft Azure or Amazon Bedrock.
Claims that Astra scored 99.9% on ARC-AGI-3, 72.6% on OSWorld 2.0 or 100% on ExploitBench are also not supported by the official material currently available.
OpenAI’s latest publicly launched model family remains GPT-5.6, comprising the Sol, Terra and Luna tiers. The company says GPT-5.6 is available across ChatGPT, Codex and the OpenAI API, with prices beginning at $5 per million input tokens and $30 per million output tokens for the flagship Sol model.
The official GPT-5.6 results differ significantly from figures attributed to Astra in the original report. OpenAI lists GPT-5.6 Sol at 62.6% on OSWorld 2.0, 73.5% on ExploitBench and 33.7% on ExploitGym under its stated evaluation conditions.
Astra’s appearance in OpenAI’s mathematical research nevertheless provides an early indication of the company’s direction. The system appears designed to extend AI beyond conventional conversation and into complex scientific work, formal reasoning and long-running research tasks.
Until OpenAI publishes a formal product announcement, however, Astra should be described as an internal research model rather than a commercially available GPT-6 release.
