How an AI Chatbot Just Solved Unsolved Math Problems for $2,000

On August 1, 2026, OpenAI published a 249-page collection of papers alongside a GitHub repository containing machine-checkable proofs that resolved ten open problems in pure mathematics and theoretical computer science. Each problem had resisted human research for at least a decade, with several standing unproven or unrefuted for over forty years.

The total token cost required for the system to generate the mathematical solutions was approximately $2,000.

The announcement sent shockwaves through the global mathematics community. The work was accomplished by an unreleased internal reasoning system codenamed Astra, designed as a multi-agent network capable of extended, test-time inference search across multi-day research tasks. Rather than delivering natural language summaries or informal sketches, OpenAI published verified Lean 4 certificates under an open-source license. In formal verification, every line of a proof must be verified by a interactive theorem prover kernel; the repository's sorry counter—the metric tracking unproven intermediate assumptions—was zero.

The findings span group theory, operator algebras, extremal combinatorics, high-dimensional geometry, lattice cryptography, and quantum complexity theory. Among the headline accomplishments, Astra constructed an explicit counterexample proving the existence of non-sofic groups—a core question in group theory open since Mikhail Gromov introduced the concept in 1999—and disproved Alain Connes's 1980 rigidity conjecture on von Neumann algebras.

The milestone represents a decisive transition in the field. For years, automated reasoning systems served primarily as proof-assistants or search engines for bounded combinatorial problems. The emergence of Astra demonstrates that scaling inference compute within formal verification environments enables AI solving math problems at the absolute frontier of academic research.

To understand how a $2,000 compute run achieved what decades of human intellectual effort could not, it is necessary to trace the escalation of artificial intelligence in pure mathematics—from early probabilistic models to the formal proof engines of 2026.

2020–2023: The Hallucination Era and the Formal Proof Bottleneck

The path toward automated mathematical discovery was initially blocked by the fundamental architecture of early large language models (LLMs). Between 2020 and 2023, standard auto-regressive models operated strictly on next-token prediction. While impressive at software engineering and conversational tasks, they struggled with rigorous multi-step deduction.

When tasked with advanced linear algebra, abstract algebra, or real analysis, early models suffered from three systemic flaws:

Semantic Drift: In long proofs requiring twenty or thirty logical steps, minor probabilistic errors compounded, leading to completely invalid conclusions.
Illusion of Rigor: Models frequently produced fluent, authoritative-sounding text containing false premises or mathematically illegal steps—a phenomenon that made human peer review cumbersome.
Lack of Internal Feedback: Unassisted language models had no mechanism to test hypotheses mid-generation. Unlike a human mathematician working at a blackboard, the model could not back-track upon hitting a logical wall.

Mathematicians quickly dismissed LLMs as general research tools. In a widely cited 2022 assessment, Fields Medalist Terence Tao noted that while AI could generate plausible research avenues, it lacked the logical precision required to generate complete, original proofs independently.

The first turn toward a solution arrived not through larger language models alone, but through the integration of interactive theorem provers (ITPs). Languages such as Lean, Coq, and Isabelle/HOL act as absolute arbiters of mathematical truth. In Lean 4, a proof is written as a sequence of formal tactics that manipulate logical statements. If a single step violates the rules of formal logic, the compiler throws an error.

The open challenge was bridge-building: traditional LLMs spoke natural language, whereas Lean demanded strict formal code.

In late 2023, early hybrid experiments began to yield results. Google DeepMind released FunSearch, a system combining a pre-trained LLM with an automated evaluator to discover new solutions to the cap set problem in extremal combinatorics. FunSearch did not solve human-posed conjectures from first principles, but it demonstrated that coupling language generation with deterministic evaluation could surface mathematical constructions superior to those found by human intuition alone.

The bottleneck, however, remained formalization. Human mathematicians rarely write proofs in Lean; translating open research problems into formal logic required months of manual labor by specialized computer scientists. Until AI could bridge natural language intuition with formal machine verification, frontier mathematics remained out of reach.

2024–2025: Test-Time Compute and the Formal Breakthroughs

The turning point occurred in mid-2024, driven by two simultaneous breakthroughs: automated auto-formalization and the scaling of inference-time search.

In July 2024, Google DeepMind unveiled AlphaProof and AlphaGeometry 2, which competed in the International Mathematical Olympiad (IMO). AlphaProof coupled a specialized language model with the Lean theorem prover, using reinforcement learning to search through millions of formal proof paths. The combined system solved four out of six problems at the IMO, achieving a silver-medal standard.

AlphaProof established a critical principle: mathematical reasoning scales with inference compute. Rather than relying solely on pre-trained parametric memory (System 1 thinking), the system spent minutes or hours exploring proof trees at test time (System 2 thinking).

                  [ Research Conjecture / Prompt ]
                                  │
                                  ▼
                     [ Natural Language Reasoning ]
                     (Multi-Agent Decomposition)
                                  │
                                  ▼
                   [ Formalization Engine (Lean 4) ]
                                  │
         ┌────────────────────────┴────────────────────────┐
         ▼                                                 ▼
[ Valid Syntax / Tactic ]                       [ Logical Error Detected ]
         │                                                 │
         ▼                                                 ▼
[ Compiler Check Pass ]                         [ Backtrack & Search Tree ]
         │                                                 │
         └────────────────────────┬────────────────────────┘
                                  │
                                  ▼
                    [ Zero-Sorry Verified Proof ]

Throughout 2025, frontier AI laboratories refocused their architectures around test-time compute. Instead of spending tens of millions of dollars solely on pre-training larger models, researchers discovered that allocating substantial compute to search during inference yielded dramatic improvements in symbolic logic, coding, and spatial reasoning.

At the same time, open-source communities accelerated the formalization of mathematics. The Lean Mathlib library expanded past two million lines of code, covering deep territories of algebraic geometry, algebraic topology, and functional analysis. This expanded the "search space" available for reinforcement learning agents.

By late 2025, the stage was set. AI models were no longer merely translating known proofs; they were learning to navigate high-dimensional proof trees independently. The remaining question was whether these systems could move beyond competition-style problems—where solutions were known to exist—and resolve genuine, open research problems where human intuition had stalled.

May 2026: The First Crack in Pure Mathematics

The preliminary confirmation came quietly on May 20, 2026. OpenAI published a paper disclosing that an internal, general-purpose reasoning system had disproved Paul Erdős’s Planar Unit Distance Conjecture.

Posed by Erdős in 1946, the conjecture asks a foundational question in discrete geometry: given $n$ points placed in a two-dimensional plane, what is the maximum number of pairs of points that can be separated by an exact distance of one unit?

For nearly eighty years, the prevailing consensus among geometricians was that optimal configurations closely mirrored regular square grids, yielding an upper bound dominated by $O(n^{4/3})$.

OpenAI’s model constructed an explicit, infinite family of point configurations that defied this eighty-year consensus. The construction drew from advanced algebraic number theory, utilizing infinite class field towers and Golod-Shafarevich theory to build point sets with significantly higher density. Princeton mathematician Will Sawin refined the model's construction, proving it yielded at least $n^{1 + 0.014}$ unit-distance pairs for infinitely many values of $n$.

The math community’s reaction shifted from skepticism to intense scrutiny. OpenAI assembled a panel of world-leading mathematicians—including Fields Medalist Sir Timothy Gowers, Noga Alon, Arul Shankar, and Jacob Tsimerman—to evaluate the output.

Gowers publicly declared that he would have recommended the proof for publication in a top-tier journal without hesitation.

The disproof of Erdős's unit distance conjecture proved that AI models were not simply executing broad combinatorial searches; they were synthesizing deep, domain-spanning theoretical frameworks to solve problems that human specialists had abandoned.

Yet, the May 2026 announcement was merely a prelude. The proof had required extensive human intervention to clean up and verify manually in natural language. The process was slow, labor-intensive, and relied on human authority for validation.

OpenAI set out to automate the entire loop.

July 2026: The Astra Demonstrations in Washington

Behind closed doors in late July 2026, the strategic focus shifted from academic validation to policy and regulatory positioning.

On July 26, 2026, OpenAI CEO Sam Altman held private briefings with key lawmakers, federal officials, and scientific advisors in Washington, D.C.. During these sessions, Altman demonstrated an unreleased model architecture codenamed Astra.

The primary purpose of the Washington briefing was to showcase Astra's reasoning capabilities under proposed federal evaluation frameworks. Under emerging administrative policy guidelines, frontier AI developers were encouraged to demonstrate verifiable safety and advanced synthetic reasoning before deploying systems capable of autonomous execution.

Altman framed Astra not as an assistant, but as an autonomous research engine. Rather than answering questions in seconds, Astra operated as an ensemble of specialized sub-agents. When fed a complex research conjecture, Astra spent hours decomposing the problem, attempting dozens of distinct algebraic approaches, testing edge cases against formal compilers, and auto-correcting its own logical flaws.

According to sources present at the demonstrations, OpenAI highlighted that Astra’s reasoning process was fully deterministic at the point of verification. While the generative exploration phase involved probabilistic search, the final output was a machine-verified mathematical proof.

Five days later, OpenAI made Astra's output public to the world.

August 1, 2026: The $2,000 Breakthrough Published

On Saturday, August 1, 2026, OpenAI published a 249-page manuscript collection titled "Ten Advances in Mathematics and Theoretical Computer Science". Accompanying the manuscript was a public GitHub repository (openai/ten-proofs) containing complete Lean 4 formalizations.

The announcement established that AI solving math problems had crossed a critical boundary: the model achieved major discoveries across eight distinct fields at a fraction of the cost of traditional academic research.

                     ASTRA $2,000 COMPUTE ADVANCEMENT
┌─────────────────────────────────────────────────────────────────────────┐
│ Total Token Cost: ~$2,000 (Sol API Rates)                              │
│ Total Formal Certificates: 10 Machine-Checkable Lean 4 Files            │
│ Lean Compiler Unproven Steps ('sorry' count): ZERO                      │
│ Manuscript Length: 249 Pages of Comprehensive Proof Walkthroughs        │
└─────────────────────────────────────────────────────────────────────────┘

OpenAI revealed that the total compute cost required for Astra to generate the mathematical solutions across all ten open problems was roughly $2,000 when priced at Sol API rates.

To put this figure into perspective, a single postdoctoral researcher at a university costs an institution roughly $80,000 to $120,000 per year in direct salary, benefits, and overhead. Many of the ten problems solved by Astra had consumed aggregate decades of grant funding, academic sabbaticals, and conference workshops without resolution.

The $2,000 figure instantly transformed the economics of theoretical research.

Crucially, OpenAI credited the AI model itself as the primary author of the mathematical content in the published paper, noting that claiming human authorship "would misrepresent both the system's contribution and the nature of genuine human intellectual work". Human researchers assisted primarily in prompting the model, organizing the natural language walkthroughs, and setting up the initial formalization harness.

Deep Dive: The Ten Frontier Problems Solved by Astra

The ten problems resolved by Astra were not contest-style puzzles; they were deep, structural roadblocks in pure and applied mathematics. Below is a detailed technical analysis of the major breakthroughs contained in the August 1 release:

1. The Existence of Non-Sofic Groups

Field: Group Theory / Symbolic Dynamics
Origin: Posed by Mikhail Gromov in 1999.
Background: A group is a mathematical structure defining symmetry. Gromov introduced the concept of "sofic groups"—groups whose global structure can be locally approximated by finite symmetric groups (analogous to approximating complex continuous systems using finite permutations). For 27 years, every known group examined by mathematicians turned out to be sofic. A fundamental question remained: do non-sofic groups actually exist, or is soficity an universal property of all groups?
Astra's Breakthrough: Astra constructed an explicit, finite presentation of a non-sofic group. By combining deep combinatorial group theory with novel topological invariants, Astra generated a group whose local neighborhood properties provably violate the finite approximation requirements of soficity.

2. Disproof of Connes's Rigidity Conjecture

Field: Operator Algebras / Functional Analysis
Origin: Posed by Fields Medalist Alain Connes in 1980.
Background: Connes’s conjecture concerned von Neumann algebras—infinite-dimensional algebras of bounded operators on a Hilbert space. Connes conjectured that for a specific rigid class of groups possessing Kazhdan's Property (T), the resulting von Neumann algebra acts as a unique fingerprint, completely determining the underlying group up to isomorphism.
Astra's Breakthrough: Astra disproved the conjecture by constructing an infinite family of non-isomorphic groups possessing Property (T) that produce identical, isomorphic von Neumann algebras. The model established that von Neumann algebras lose certain structural details regarding group isomorphism, refuting a 46-year-old foundational hypothesis in operator theory.

3. Proof of Ehrhart's Volume Conjecture

Field: Lattice Geometry / Convex Polyhedra
Origin: Discrete geometry literature on lattice polytopes.
Background: Ehrhart polynomials count the number of integer lattice points inside scaled versions of convex polytopes. Ehrhart hypothesized a strict inequality bounding the geometric volume of certain high-dimensional lattice polytopes relative to their degree coefficient structures.
Astra's Breakthrough: Astra provided a complete, rigorous proof establishing the upper bounds on polytope volumes across arbitrary dimensions, validating Ehrhart's volume conjecture and providing new tools for algorithmic integer programming.

4. Resolution of Erdős Problem #183

Field: Extremal Combinatorics / Ramsey Theory
Origin: Paul Erdős's catalog of open problems.
Background: Erdős Problem 183 asks for tight asymptotic bounds on multicolor Ramsey numbers—specifically, determining the minimum size of a complete graph required to guarantee the existence of monochromatic triangles when edges are colored using $k$ distinct colors.
Astra's Breakthrough: Astra established tighter upper and lower bounds for Erdős Problem 183, reducing the asymptotic gap that had remained static since the 1970s through a novel probabilistic graph construction.

┌────────────────────────────────────────────────────────────────────────┐
│                   SUMMARY OF HIGHLIGHTED RESULTS                       │
├───────────────────────┬───────────────────────────┬────────────────────┤
│ Problem               │ Field                     │ Status / Output    │
├───────────────────────┼───────────────────────────┼────────────────────┤
│ Non-Sofic Groups      │ Group Theory              │ Proved Existence   │
│ Connes's Rigidity     │ Operator Algebras         │ Disproved          │
│ Ehrhart's Volume      │ Lattice Geometry          │ Proved Conjecture  │
│ Erdős Problem #183    │ Ramsey Theory             │ Tighter Bounds     │
│ High-Dim Sphere Pack  │ Metric Geometry           │ Tighter Bounds     │
│ Permanent Complexity  │ Arithmetic Circuits (TCS) │ Lower Bounds       │
│ 2-Player Quantum Games│ Quantum Complexity Theory │ Repetition Theorem │
└───────────────────────┴───────────────────────────┴────────────────────┘

5. High-Dimensional Sphere Packing Bounds

Field: Metric Geometry / Coding Theory
Background: Sphere packing asks how densely identical spheres can be packed in $n$-dimensional space. While Maryna Viazovska won the Fields Medal in 2022 for solving the problem in dimensions 8 and 24, high-dimensional asymptotic density limits remained wide open.
Astra's Breakthrough: Astra constructed optimized linear programming bounds that improved the upper density limits down to the Cohn-Elkies threshold across high dimensions, providing immediate applications to error-correcting codes in telecommunications.

6–10. Complexity, Cryptography, and Quantum Games

Arithmetic Circuit Complexity: Astra established new, explicit lower bounds on the size of arithmetic circuits required to compute the permanent of a matrix, making progress on Valiant's conjecture (the algebraic equivalent of P vs. NP).
Quantum Complexity: The model proved a generalized parallel repetition theorem for two-player quantum non-local games, resolving a long-standing conjecture regarding quantum entanglement entanglement-assisted communication protocols.
Lattice Cryptography: Astra discovered novel reduction bounds on the Shortest Vector Problem (SVP) in high-dimensional lattices, refining the mathematical assumptions underpinning post-quantum cryptographic algorithms.
Coding Theory & Extremal Graph Theory: The final two papers established optimal code rates for non-linear binary codes and resolved an open question regarding extremal subgraphs in hypergraph networks.

Technical Mechanics: How Astra Solved These Problems for $2,000

Understanding why Astra succeeded where human researchers and earlier AI systems failed requires looking at the computational architecture behind the $2,000 inference run.

OpenAI researchers, led by reasoning specialist Noam Brown, leveraged a paradigm focused on multi-agent collaboration, formal feedback loops, and dynamic search allocation.

The Multi-Agent Search Loop

Rather than treating a complex proof as a single, linear generation task, Astra operates as an orchestrated network of distinct agent roles:

The Conjecturer/Architect: Formulates high-level strategies, breaks down the main conjecture into smaller sub-lemmas, and suggests abstract algebraic approaches.
The Formalizer: Translates natural language mathematical arguments into syntactically valid Lean 4 tactic strings.
The Verification Kernel: A deterministic, non-neural compiler (the Lean 4 kernel) that evaluates each tactic. If a step fails, the compiler returns precise error locations to the model.
The Refutation Agent: Actively searches for counterexamples or invalid logical steps, attempting to break proposed sub-lemmas before full formalization begins.

       [ Research Problem Input ]
                   │
                   ▼
       [ Conjecturer / Architect Agent ] ──► (Decomposes into Lemmas)
                   │
                   ▼
       [ Refutation Agent (Stress Test) ] ──► (Searches Counterexamples)
                   │
                   ▼
       [ Formalization Agent (Lean 4) ] ──► (Translates to Tactics)
                   │
                   ▼
       [ Deterministic Lean Kernel ] ──► Pass: [ Zero-Sorry Proof ]
                   │
              Fail: Error Logs sent back to Search Tree

Eliminating the "Hallucination Tax"

In standard language model runs, a hallucination ruins the downstream output, requiring expensive human intervention to spot errors. By routing all output through the Lean 4 proof environment, Astra eliminated the hallucination tax.

If Astra proposed an illegal step, the Lean compiler rejected it within milliseconds. The system used Monte Carlo Tree Search (MCTS) combined with reinforcement learning to backtrack, explore alternate proof paths, and learn which tactic sequences successfully moved the proof state closer to completion.

Test-Time Compute Efficiency

The $2,000 cost figure reflects optimized token consumption. Instead of running massive parallel compute clusters blindly, Astra dynamically adjusted its inference budget based on search complexity:

For routine algebraic manipulations, the system used lightweight, high-speed token generation.
When encountering critical structural bottlenecks (such as constructing the non-sofic group presentation), Astra scaled up deep search trees, exploring millions of potential tactic paths over several days before synthesizing the correct proof strategy.

This targeted distribution of compute allowed Astra to solve decade-old research problems for the price of a mid-tier developer laptop.

Reactions from the Mathematical Community

The publication of ten machine-verified proofs triggered immediate reaction across academia.

Thomas Bloom, a mathematician at the University of Manchester who maintains the global catalog of Erdős problems, described the results as "big news" and a landmark moment for structural combinatorics. Bloom noted that Astra’s results went far beyond simple automated searches, demonstrating a genuine capacity to discover non-trivial mathematical constructions.

Fields Medalist Sir Timothy Gowers remarked on the clarity and verification quality of the outputs, emphasizing that the presence of machine-checkable Lean certificates completely changes the nature of peer review.

Traditionally, verifying a 50-page pure mathematics manuscript requires months or years of painstaking effort by expert peer reviewers—a process historically vulnerable to overlooked errors. With Lean 4 certificates, verification takes minutes and carries complete mathematical certainty.

┌─────────────────────────────────────────────────────────────────────────┐
│                    PERSPECTIVES FROM THE FIELD                          │
├───────────────────┬─────────────────────────────────────────────────────┤
│ Expert            │ Reaction & Commentary                               │
├───────────────────┼─────────────────────────────────────────────────────┤
│ Sir Timothy Gowers│ "Would recommend for a top journal without          │
│ (Fields Medalist) │ hesitation. The Lean certificates remove ambiguity."│
├───────────────────┼─────────────────────────────────────────────────────┤
│ Thomas Bloom      │ "Big news... even more significant than the May     │
│ (Univ. Manchester)│ unit-distance disproof. A major step for research." │
├───────────────────┼─────────────────────────────────────────────────────┤
│ Noam Brown        │ "Sadly, no Millennium Prize Problems (yet)... but   │
│ (OpenAI Lead)     │ test-time compute has immense runway ahead."        │
└───────────────────┴─────────────────────────────────────────────────────┘

However, researchers also emphasized important caveats.

OpenAI co-author Noam Brown noted on X that while Astra's achievements were extraordinary, the system had also attempted—and failed—to solve several other famous open problems, including the $1 million Millennium Prize Problems managed by the Clay Mathematics Institute.

"Sadly, no Millennium Prize Problems (yet)," Brown wrote. "We didn't spend a lot on each problem, and it is possible to push test-time compute much further. Astra is a major step for scientific reasoning."

Other critics pointed out that the $2,000 figure represents the cost of the successful token generation runs, rather than the total historical compute spent training the underlying base model or exploring abandoned proof attempts. Furthermore, because Astra remains an internal, unreleased model, independent researchers cannot yet re-run the generative search process itself—they can only verify the final Lean outputs published on GitHub.

Analysis: The Changing Economics of Mathematical Discovery

The emergence of AI solving math problems at this level fundamentally alters the relationship between compute, funding, and scientific progress.

Historically, progress in pure mathematics has been constrained by human capital. The number of active researchers capable of understanding deep operator algebras or high-dimensional geometry is small, and training a new PhD mathematician requires nearly a decade of higher education.

Astra introduces an era where compute can directly substitute for scarce cognitive throughput in specific symbolic domains.

TRADITIONAL DISCOVERY MODEL           ASTRA COMPUTE DISCOVERY MODEL
┌───────────────────────────┐         ┌───────────────────────────┐
│ Years of Human Effort     │         │ ~$2,000 Inference Run     │
│ Grant Funding & Overhead  │   ──►   │ Multi-Agent MCTS Search   │
│ Human Peer Review (Years) │         │ Instant Lean Verification │
└───────────────────────────┘         └───────────────────────────┘

When the cost of generating a novel theoretical proof drops to $2,000, the bottleneck shifts:

From Proof Generation to Problem Formulation: The high-value skill for human mathematicians will increasingly shift toward posing the right conjectures, defining novel conceptual frameworks, and identifying which problems are susceptible to formal search.
The Death of Unverified Literature: As AI systems scale, unverified natural language manuscripts may become obsolete in technical subfields. Future journal submissions may routinely require machine-checkable certificates alongside human-readable explanations.
Cross-Domain Synthesis: AI models trained across all mathematical disciplines simultaneously can draw connections between distant fields—such as linking algebraic class field towers to geometric unit-distance problems—in ways that hyper-specialized human academics rarely attempt.

What to Watch Next

As the mathematics and AI communities process the August 1 release, several critical milestones will dictate the next phase of development:

Independent Lean Audits: Academic teams are currently inspecting the openai/ten-proofs repository to evaluate the elegance, efficiency, and human-readability of the AI-generated tactic code.
Scaling Up Test-Time Budgets: OpenAI’s $2,000 run was relatively modest. Research labs will soon experiment with allocating $100,000 or $1,000,000 in test-time compute to a single, high-value problem. Whether this order-of-magnitude increase in search scale can crack a Millennium Prize Problem like the Riemann Hypothesis or Navier-Stokes existence remains the central open question.
Commercial Release and GPT-6 Integration: OpenAI has not yet announced a formal release date for Astra, nor whether it will be integrated into ChatGPT under a GPT-6 designation or launched as a specialized enterprise scientific API.
Competitive Responses: Competitors including Google DeepMind, Anthropic, and open-source theorem-proving projects are preparing rival reasoning models designed specifically for deep scientific discovery.

The disproofs of Connes's conjecture and the construction of non-sofic groups demonstrate that artificial intelligence has crossed from an assistant tool into an active collaborator at the edge of human knowledge. For $2,000, the baseline for scientific discovery was permanently reset.