Anthropic Google Research Race Meets OpenAI's Ten-Proof Claim

OpenAI entered the Anthropic Google research race with ten claimed advances across mathematics and theoretical computer science, all generated by an unreleased model called Astra.

The company says these were not benchmark exercises or familiar olympiad questions. Each problem had remained open, without progress on its main result, for at least a decade. Several had resisted mathematicians for much longer.

That distinction turns the announcement into something more consequential than another model score. OpenAI is claiming that one system produced original arguments across ten specialized fields, then translated those arguments into machine-checkable Lean 4 certificates.

The timing also matters. Anthropic recently described Claude Mythos Preview finding new attacks against cryptographic constructions. Google DeepMind has spent years moving from olympiad proofs toward tools for mathematical research.

OpenAI is now making the broadest research claim of the three. It says Astra generated results involving geometry, coding theory, group theory, quantum complexity, cryptography, and combinatorics.

Yet the announcement also exposes a crucial divide. A Lean certificate can verify that formal statements follow from encoded assumptions. It cannot establish that those statements capture every intended theorem, historical dependency, or disciplinary nuance.

The ten results therefore begin a review process rather than end one. Mathematicians must inspect the manuscripts, definitions, novelty claims, and relationship to earlier work.

That burden is substantial. If even several results survive expert scrutiny, the Anthropic Google competition will no longer center only on coding, benchmarks, and scientific assistance. It will include the production of new research.

OpenAI Says Astra Produced Ten Long-Stalled Results

OpenAI's central claim is that Astra crossed from solving supplied problems into generating publishable mathematical research.

The company announced the work on August 1, 2026. Its accompanying research publication describes ten problems spanning pure mathematics and theoretical computer science.

OpenAI says an internal version of Astra found the mathematical arguments. Humans then prepared those arguments as manuscripts with help from the same model.

Astra subsequently formalized each result in Lean 4. Lean is a proof assistant that checks whether a formal argument follows from explicitly stated definitions and rules.

OpenAI released those certificates through the public ten-proofs repository. The repository includes separate Lean files for every claimed result and instructions for building them with mathlib.

The first result concerns high-dimensional sphere packing. This field asks how densely identical spheres can fit in spaces whose dimensions extend far beyond ordinary geometry.

OpenAI claims an improved general upper bound that reaches the Cohn-Elkies threshold. Its paper describes this as the first improvement to the general asymptotic sphere-packing exponent since 1978.

The second result concerns binary and spherical codes. Such codes represent arrangements whose minimum separation helps determine how reliably information can survive noise.

Astra reportedly found exponentially stronger upper bounds for binary codes at every prescribed minimum distance. The company says related spherical-code bounds recover the new sphere-packing exponent.

The third result constructs a non-sofic group. Sofic groups admit approximations through finite permutation structures, and mathematicians had long asked whether every countable group possessed that property.

The claimed construction gives a negative answer. If accepted, it resolves a central question connecting group theory, dynamics, and operator algebras.

The fourth result presents counterexamples to Connes's rigidity conjecture. OpenAI says Astra constructed infinitely many nonisomorphic groups sharing the same group von Neumann algebra.

The fifth result targets arithmetic circuit complexity. It gives new lower bounds for computing the permanent, a matrix function resembling the determinant but lacking its simplifying signs.

Lower bounds matter because they identify tasks that restricted computational systems cannot perform efficiently. Strong unconditional results remain unusually difficult in complexity theory.

The sixth result claims exponential parallel repetition for every finite two-player quantum game. Parallel repetition studies how repeatedly playing a game changes the probability of winning all copies.

The seventh addresses the closest vector problem. It asks for the lattice point nearest a given target and underpins parts of computational complexity and lattice cryptography.

OpenAI reports polynomial-factor hardness for approximating the Euclidean version. The paper also claims consequences for decoding and related lattice problems.

The eighth result concerns Ehrhart's volume conjecture. It identifies the sharp maximum volume for certain convex bodies containing only their centroid as an interior lattice point.

The ninth gives a superexponential lower bound for multicolor triangle Ramsey numbers. OpenAI says this resolves problem 183 from a collection associated with Paul Erdős.

The final result supplies counterexamples to two extremal graph theory conjectures. Those constructions reportedly resolve Erdős problems 146 and 180.

This range is itself part of the claim. Astra did not specialize in one narrow formal system or a single mathematical domain.

OpenAI says the problems saw no progress on their main results for at least ten years. That selection criterion attempts to separate genuine research from routine theorem completion.

Still, "no progress on the main result" requires careful interpretation. A field can accumulate techniques, partial results, and adjacent insights without settling its headline question.

Those developments may supply essential ingredients for a later solution. Establishing Astra's contribution therefore requires more than confirming the final proof term.

It requires reconstructing how every argument relates to the literature. That is where the announcement moves from an engineering demonstration into academic review.

Why the Anthropic Google Race Now Includes Original Research

The competitive target has shifted from answering hard questions to selecting, solving, and certifying questions without a human-authored route.

Google DeepMind established an important earlier reference point. AlphaGeometry combined a language model with symbolic deduction to solve difficult geometry problems through synthetic training data.

AlphaProof later used reinforcement learning and Lean to attack International Mathematical Olympiad problems. Google reported that AlphaProof and AlphaGeometry 2 together reached silver-medal performance at the 2024 competition.

Olympiad problems are demanding, but they arrive with known solutions and carefully bounded statements. Open research presents a different challenge because neither the route nor the answer is known.

Google has since pushed toward that broader target. Its AI for Math initiative connects researchers with Gemini Deep Think, AlphaEvolve, and AlphaProof.

AlphaEvolve searches for algorithms by combining language models with automated evaluators. AlphaProof focuses on formal proof completion, where a checker can supply a clear correctness signal.

That Anthropic Google landscape now has a third model of research automation. OpenAI's Astra claim combines open-ended argument generation with formal certification across unrelated specialties.

Anthropic's recent cryptography work offers the closest comparison. The company says Claude Mythos Preview found improved attacks against two carefully studied cryptographic constructions.

One involved HAWK, a digital signature design considered in post-quantum cryptography. Anthropic says the model found a previously unknown attack after extended autonomous work.

The other involved a reduced-round form of AES, meaning a research variant with fewer transformation rounds than production AES. Mythos reportedly accelerated an existing attack against that variant.

Anthropic explicitly noted that neither result broke deployed encryption. The importance rests in the model contributing novel cryptanalytic ideas, not in compromising current systems.

Its cryptography assessment also illustrates how frontier laboratories are changing their evaluation methods. They increasingly test models on unresolved, professionally meaningful tasks.

Traditional benchmarks become less informative once their questions enter training sets or public solution archives. Open problems offer fresher evidence, but they also weaken standardized comparison.

A laboratory chooses the problems, allocates computational effort, and decides which successes to publish. Outsiders rarely know how many attempted problems produced nothing useful.

That denominator matters. Ten successes from ten attempts would imply something different from ten successes selected after a much larger search.

OpenAI has published detailed outputs for the successful cases. It has not provided a complete accounting of Astra's failed research attempts or its broader selection process.

This limits direct comparison among Astra, Mythos, and Google's systems. Each company uses different domains, tools, budgets, supervision, and definitions of success.

Even so, their direction is aligned. Frontier laboratories are treating research discovery as both a product capability and a model evaluation.

The resulting pressure extends beyond the three companies. Specialized theorem-proving startups must show whether their systems offer better reliability, access, or researcher control.

Universities also face pressure. Their researchers need enough access and infrastructure to reproduce claims produced inside private laboratories.

Publishers and conferences must decide how to review AI-generated work. Conventional peer review assumes that human authors can explain choices, answer objections, and accept intellectual responsibility.

OpenAI's announcement challenges that assumption. The company says the arguments came from Astra while humans prepared and formalized them.

That attribution is unusually direct. It avoids presenting machine-generated ideas as entirely human work, but it leaves responsibility distributed across a model and its operators.

The competitive question is no longer simply which system gets more answers right. It is which organization can make machine-generated research credible to communities that define credibility themselves.

Lean Certificates Change Verification, Not Scientific Judgment

Formal verification narrows the error surface, but it does not turn a company announcement into settled mathematics.

A conventional mathematical proof is written for expert readers. Referees examine whether its reasoning works, whether cited results apply, and whether hidden assumptions undermine the conclusion.

A Lean proof operates differently. Every step must fit a formal language, and the proof assistant's small trusted kernel checks the resulting term.

This provides a strong safeguard against many familiar errors. A model cannot simply skip a difficult implication if Lean requires a term establishing it.

The public repository also gives independent reviewers something concrete to run. Its files target Lean 4.32.0 and rely on mathlib, the community library supporting formalized mathematics.

OpenAI includes a second verification route through Comparator challenges. Comparator is intended to help check certificates while reducing trust in generated support code.

These choices make the work more auditable than a collection of model transcripts. Anyone with the necessary environment can inspect definitions and build the formalizations.

However, a compiled proof answers a precise question: does this encoded theorem follow from these encoded assumptions within this formal environment?

It does not automatically answer whether the encoded theorem matches the informal headline. That correspondence still requires mathematical judgment.

A definition can omit a condition that specialists consider essential. A formal theorem can establish a weaker statement than readers infer from its description.

Imported library results also carry context. Reviewers must understand whether Astra used established theorems appropriately and whether the formal statement preserves their intended scope.

Novelty creates another challenge. Lean does not search the historical literature and decide whether an argument appeared under different terminology decades earlier.

That task demands specialists who know the field's techniques and unpublished folklore. It may take longer than checking whether the certificate compiles.

The manuscripts must also become understandable. Mathematics advances when researchers can reuse an idea, connect it to other work, and explain why the mechanism succeeds.

A huge formal term can certify correctness without supplying that understanding. OpenAI's reasoning walkthroughs attempt to bridge the gap, but they remain model-generated narratives.

The distinction resembles verified software. A program can satisfy a formal specification while the specification fails to express what users actually need.

That does not diminish formal verification. It clarifies its role as one strong layer inside a larger system of evidence.

The ten-proof release represents a notable combination of natural-language discovery and formal checking. Previous systems often excelled at one side while struggling with the other.

Language models can generate plausible arguments across many domains. Proof assistants can reject invalid steps but demand precise representations and extensive supporting definitions.

A system that moves effectively between those modes reduces a major bottleneck. It can explore loosely, organize an argument, then translate the result into a machine-checkable object.

That workflow also changes the human role. Mathematicians may spend less time repairing local derivations and more time choosing definitions, reviewing novelty, and extracting reusable concepts.

Yet certification can create false confidence when readers do not distinguish formal validity from scientific importance. A checked theorem may be correct but minor, redundant, or framed misleadingly.

The ten results do not appear minor on their face. Several resolve named conjectures or longstanding questions, according to OpenAI.

Their actual importance will emerge through independent reading, attempted simplification, and follow-on research. A proof earns its place through community use, not compilation alone.

This is why the Anthropic Google research race cannot be scored with a single benchmark. The relevant outputs are arguments, tools, and new lines of inquiry.

Their value depends on whether outside researchers can inspect them and build upon them. Public certificates improve that possibility, but access to Astra remains closed.

The Missing Denominator Is the Biggest Reason for Caution

OpenAI has shown ten selected successes, while the scale and outcome of its unsuccessful search remain unknown.

Research productivity cannot be inferred from successful outputs alone. Readers need some view of the attempts, interventions, and evaluation criteria surrounding those outputs.

OpenAI says Astra generated the arguments and later formalized them. The publication does not fully describe how humans selected problems or redirected stalled runs.

It also does not disclose how many candidate solutions failed mathematical review. That missing denominator makes efficiency claims difficult to interpret.

A system can appear extraordinarily productive if only its strongest results become visible. Scientific evaluation normally guards against this through preregistration, repeated experiments, and independent replication.

Open-ended mathematics does not fit those methods neatly. Problems vary dramatically, and a single idea can matter more than hundreds of failed attempts.

Still, laboratories can provide fuller records. They can publish problem-selection rules, unsuccessful runs, intervention logs, and fixed evaluation windows.

Those materials would help outsiders distinguish general research ability from a carefully curated portfolio. They would also reveal where human judgment entered the process.

Human involvement is not a flaw. Mathematical research has always been collaborative, and tools routinely shape which routes researchers explore.

The issue is attribution. Readers need enough detail to understand whether Astra originated the decisive ideas or completed a structure supplied by expert operators.

OpenAI takes an unusually clear position here. It says human authorship would misrepresent proofs whose mathematical arguments were generated entirely by its system.

That stance responds directly to a concern raised by the Leiden Declaration. The declaration calls for transparency, proper attribution, reproducibility, and continuing human responsibility.

It also warns that automated techniques can produce plausible but unreliable arguments. Formal certificates address part of that concern, but not every institutional consequence.

Private models may influence which problems receive attention. Laboratories could prioritize fields with clean automated feedback while neglecting research requiring experiments, interpretation, or social context.

Access creates another tension. A closed model can produce public proofs, yet researchers outside its operator's network cannot probe its behavior under comparable conditions.

Google and Anthropic face the same pressure. Their most capable research systems also depend on private infrastructure, proprietary training methods, and controlled access.

That arrangement concentrates agenda-setting power. A small number of companies can choose which scientific problems receive vast computational resources.

It can also widen inequality among universities. Researchers at wealthy institutions may secure partnerships while others work only with public outputs.

The history of computer-assisted proof offers a useful precedent. The four-color theorem initially drew skepticism because its verification required extensive computation beyond ordinary hand checking.

Over time, clearer methods and independently checkable implementations strengthened acceptance. The debate helped mathematics refine its standards for computational evidence.

Astra's results present a related problem at a larger scale. The machine is not only checking cases; it is reportedly generating central conceptual arguments.

The appropriate response is neither automatic rejection nor immediate acceptance. It is deeper disclosure, independent verification, and explicit responsibility.

The repository is a meaningful start because it exposes formal artifacts. The paper and walkthroughs provide additional material for specialists.

However, the strongest test will come from researchers unaffiliated with OpenAI. They must confirm the statements, compare prior literature, and explain the arguments in their own terms.

Errors may still emerge. Some claims may require narrower wording, additional hypotheses, or revised novelty descriptions.

Such corrections would not necessarily invalidate Astra's research ability. Human papers also change during peer review, sometimes substantially.

The critical measure is whether the process converges efficiently toward reliable knowledge. A system that generates many plausible dead ends could overwhelm scarce reviewing capacity.

That risk is already visible across research publishing. Models can produce technical prose faster than experts can assess it.

Formalization helps by filtering invalid deductions. It does not solve literature review, significance assessment, or the limited supply of specialist attention.

The next research benchmark may therefore be institutional rather than mathematical. Frontier labs must show they can generate results without flooding communities with unreviewable claims.

What the Ten Proofs Put Under Pressure

Astra places pressure on research workflows, specialist AI systems, and the assumption that original mathematics requires a human-generated proof strategy.

For developers, the immediate lesson concerns architecture. The important capability is not isolated text generation but a loop connecting exploration, tools, formal feedback, and revision.

Lean supplies an evaluator with unusually clear signals. A candidate proof either passes the kernel or produces an error that guides another attempt.

Comparable evaluators exist in software development, circuit design, and some scientific simulations. Fields without reliable feedback remain harder to automate.

This helps explain why mathematics has become a major frontier for agentic systems. It offers deep intellectual problems alongside increasingly mature verification tools.

The results also pressure specialized theorem provers. A broad model that works across ten fields can compete with systems trained for narrower formal tasks.

Specialists retain advantages, including lower operating requirements, reproducibility, and tighter integration with mathematical libraries. Open systems can also support independent experimentation.

The winning approach may combine them. A broad model can propose strategies while a specialized prover searches locally and checks every step.

For mathematicians, the change is less about immediate replacement than workflow reallocation. Literature search, lemma discovery, proof repair, and formal translation all become candidates for automation.

The most valuable human contribution may shift toward problem selection and conceptual interpretation. Researchers will still decide which statements matter and which abstractions illuminate them.

Those decisions are not cosmetic. A technically correct theorem can remain sterile until someone sees how it connects to a wider body of ideas.

Enterprises should also notice the verification pattern. Astra's announcement suggests that frontier models become more useful when their output must pass an external checker.

That principle applies beyond mathematics. Code can face tests, database queries can face schemas, and analytical claims can face reproducible calculations.

Knowledge workers cannot formalize every conclusion. They can still preserve sources, assumptions, and decision trails so another person can audit the result.

A searchable AI knowledge base becomes more relevant as models generate larger volumes of candidate research. Verification depends on retaining the evidence behind each conclusion.

The pressure also reaches peer review. Journals may need reviewers who can read both informal manuscripts and proof-assistant code.

Formal artifacts could eventually reduce some checking work. Initially, they may increase it because reviewers must compare two representations of every result.

Academic credit policies will need similar revision. A paper can involve model-generated ideas, human curation, formalization, engineering, and independent mathematical explanation.

Calling every contributor an author may blur meaningful differences. Excluding human operators may also hide essential intellectual and technical labor.

OpenAI's explicit attribution gives institutions a concrete case to debate. The company accepts responsibility while declining to claim that humans created Astra's mathematical arguments.

That framework will face tests if errors appear. Responsibility must include correction, documentation, and support for independent examination.

The Anthropic Google comparison sharpens this issue because research outputs can create external risks. Cryptographic discoveries demand coordinated disclosure before full publication.

Pure mathematics usually creates fewer immediate security concerns. The closest vector result, however, touches a field closely related to post-quantum cryptography.

A theorem about computational hardness can affect security assumptions even without revealing a deployed vulnerability. Specialists must interpret exactly what the reduction establishes.

This interaction between mathematical discovery and operational risk will grow. Models that move across disciplines can find results whose consequences exceed the original evaluation.

Laboratories need review processes that recognize those connections. A mathematically valid output may require security, policy, or domain-specific assessment before release.

Three Signals Will Decide Whether Astra Changed Mathematics

The next stage depends on independent validation, reproducible model access, and evidence that the ten proofs generate further human research.

The first signal is specialist review of the ten manuscripts and Lean certificates. Researchers must confirm that each formal statement matches its advertised result.

Watch for public explanations from experts in sphere packing, group theory, quantum complexity, and extremal combinatorics. Their assessments will carry more weight than general AI commentary.

If several communities accept the results with only minor corrections, OpenAI's central claim becomes much stronger. Major gaps or narrowed statements would weaken it.

The second signal is whether OpenAI enables reproducible evaluation of Astra or a closely related model. Public outputs alone cannot establish how reliably the system performs research.

A credible evaluation would use previously unseen problems, fixed conditions, and independent judges. It should report unsuccessful attempts alongside successes.

Google and Anthropic will face the same expectation. The Anthropic Google race becomes scientifically useful only when outsiders can compare systems under meaningful conditions.

The third signal is follow-on mathematics. Important proofs usually introduce techniques that other researchers adapt, simplify, or apply elsewhere.

OpenAI notes that an earlier AI-generated disproof of the Erdős unit-distance conjecture already inspired related work. The ten new results now face that same practical test.

If researchers extract reusable ideas, the announcement will represent more than accelerated proof production. It will show that a model can contribute to the evolving language of mathematics.

If the arguments remain opaque formal artifacts, their influence will be narrower. Correctness matters, but understanding determines how far a result travels.

The coming months should also reveal whether other laboratories answer with comparable open-problem portfolios. Google has the formal systems, while Anthropic has shown extended autonomous research in cryptography.

A response does not need another list of ten results. A smaller number of independently validated discoveries could provide stronger evidence.

Readers should resist treating this as a corporate medal table. Mathematics is not won by accumulating the most conjectures in one announcement.

The deeper contest concerns trustworthy research infrastructure. Systems must generate ideas, expose evidence, accept correction, and leave experts able to understand what changed.

OpenAI has supplied unusually concrete artifacts for that contest. The paper, walkthroughs, and Lean repository give specialists more than a press release to examine.

It has not supplied the full experimental denominator or broad access to Astra. Those gaps prevent a final judgment about the model's general research ability.

The Anthropic Google research race has still crossed a threshold. Frontier laboratories now present original scientific results as evidence of model capability, not merely as downstream applications.

That changes what developers, researchers, and institutions should ask. The key question is no longer whether AI can produce a convincing proof.

It is whether independent experts can verify the statement, understand the idea, reproduce the process, and build something valuable from it.

Follow the certificates rather than the announcement. Then watch what mathematicians do with them.