{"slug": "openai-says-astra-produced-ten-mathematics-and-computer-science-results", "title": "OpenAI Says Astra Produced Ten Mathematics and Computer Science Results", "summary": "OpenAI disclosed on August 1 that an internal version of Astra, described as its next major model, produced ten new results in mathematics and theoretical computer science, with problems that had seen no progress for at least a decade. The company released human-prepared manuscripts, Lean formalizations, and model reasoning narrations, estimating the total token cost at roughly $2,000 at Sol API rates.", "body_md": "# OpenAI Says Astra Produced Ten Mathematics and Computer Science Results\n\nOpenAI disclosed on August 1 that an internal version of Astra, described as its next major model, produced ten results in mathematics and theoretical computer science. OpenAI reported that the problems had seen no progress on their main results for at least a decade, and released human-prepared manuscripts, Lean formalizations, and model reasoning narrations for the work.\n\nOpenAI disclosed on August 1 that an internal version of **Astra**, which it describes as its next major model, produced ten new results in mathematics and theoretical computer science. In a post titled \"Ten advances in mathematics and theoretical computer science,\" OpenAI reported that the selected problems had seen no progress on their main results for at least a decade, and in most cases considerably longer.\n\nThe company stated that it is releasing manuscripts prepared by humans with assistance from the same model, a formalization of each argument in **Lean**, and a model narration of the reasoning process for each solution. OpenAI estimated that the total token cost of finding the ten solutions would be roughly **$2,000 at Sol API rates**.\n\n### Ten reported results across mathematics and theory\n\nOpenAI's list spans high-dimensional geometry, coding theory, group theory, operator algebras, arithmetic circuit complexity, quantum complexity, lattice cryptography, and extremal combinatorics. The reported results include:\n\n- •New upper bounds on high-dimensional sphere-packing density.\n- •Improved bounds for binary and spherical codes.\n- •A construction establishing the existence of non-sofic groups.\n- •A disproof of Connes's rigidity conjecture concerning groups and von Neumann algebras.\n- •New lower bounds for arithmetic circuits computing the permanent.\n\nThe Next Web reports that the package also includes results concerning Ehrhart's volume conjecture, parallel repetition for two-player quantum games, and three problems associated with Paul Erdos. It describes the non-sofic-group construction as the first explicit result of its kind, addressing a question associated with soficity that had remained unresolved since the concept was introduced in 1999.\n\nSebastien Bubeck, OpenAI's head of mathematics research, wrote on X that the results were \"beautiful\" and that each release included a Lean certificate and a chain-of-thought walkthrough, according to The Next Web. OpenAI's original post refers instead to a \"model's narration of its thinking process.\"\n\n### Formal artifacts make the claims more inspectable\n\nThe release is notable not only for the asserted mathematical discoveries but also for the accompanying proof artifacts. Lean is an interactive theorem prover: a formal certificate can be checked by software against explicitly encoded definitions, axioms, and inference rules. That does not by itself determine whether a result is mathematically important or whether its informal framing is complete, but it gives researchers a concrete object for independently inspecting formalized arguments.\n\nFor ML practitioners, this is a more consequential evaluation format than reporting a score on a benchmark alone. Open research problems require a system to generate candidates that withstand expert and formal scrutiny, while theorem-proving workflows also expose practical constraints around translation from natural-language arguments into a proof assistant.\n\n### What remains undisclosed\n\nOpenAI did not publish a consumer-facing release date, model card, architecture, context-window specification, or API details for Astra in the August 1 post. Gizmodo reported on August 2 that OpenAI had not replied to its questions about Astra's official name or its relationship to other unreleased models mentioned in prior company material.\n\nGizmodo also cited an anonymously sourced report claiming Astra can perform \"long-running\" work. That capability claim has not been substantiated in OpenAI's mathematics post, which focuses on the reported research outputs and associated formalizations.\n\nThe release follows OpenAI's May announcement of a result related to the Erdos unit-distance conjecture, which its August post described as having been discovered while evaluating an unreleased model. The present set of ten results broadens that public evidence from a single problem area to several domains, but the practical availability and general-purpose behavior of Astra remain unannounced.\n\n## Key Points\n\n- 1OpenAI identified Astra as its next major model while publishing ten reported advances across mathematics and theoretical computer science.\n- 2Lean formalizations provide inspectable proof artifacts, enabling stronger independent verification than benchmark scores or informal model demonstrations alone.\n- 3The release includes human-prepared manuscripts, Lean formalizations, and model narrations for each of the ten solutions.\n\n## Scoring Rationale\n\nThe disclosure identifies an unreleased major OpenAI model and attaches it to ten reported research-level mathematical results with formal Lean artifacts. The work is highly relevant to researchers building reasoning and theorem-proving systems, although Astra's availability, specifications, and broader performance remain undisclosed.\n\n## Sources\n\nPrimary source and supporting public references used for this report.\n\nPractice interview problems based on real data\n\n1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.\n\n[Try 250 free problems](/problems)", "url": "https://wpnews.pro/news/openai-says-astra-produced-ten-mathematics-and-computer-science-results", "canonical_source": "https://letsdatascience.com/news/openai-reveals-astra-through-mathematics-results-731984be", "published_at": "2026-08-02 09:00:45+00:00", "updated_at": "2026-08-02 10:58:49.413840+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-products"], "entities": ["OpenAI", "Astra", "Lean", "Sebastien Bubeck", "The Next Web", "Gizmodo", "Paul Erdos", "Connes"], "alternates": {"html": "https://wpnews.pro/news/openai-says-astra-produced-ten-mathematics-and-computer-science-results", "markdown": "https://wpnews.pro/news/openai-says-astra-produced-ten-mathematics-and-computer-science-results.md", "text": "https://wpnews.pro/news/openai-says-astra-produced-ten-mathematics-and-computer-science-results.txt", "jsonld": "https://wpnews.pro/news/openai-says-astra-produced-ten-mathematics-and-computer-science-results.jsonld"}}